INFORMATION ENRICHMENT FOR QUALITY RECOMMENDER SYSTEMS · Recommender systems have been an active research area for more than a decade. Many different techniques and systems with

INFORMATION ENRICHMENT FOR

QUALITY RECOMMENDER SYSTEMS

Li-Tung Weng (B.Sc. (Hons))

A Dissertation

Submitted in Fulfil of the Requirements for the Degree of

Doctor of Philosophy

Faculty of Information Technology

Queensland University of Technology

Brisbane, Australia

November 2008

Page i

Keywords

Collaborative Filtering, Cold-Start Problem, Distributed Systems, Ecommerce, Product

Taxonomy, Recommendation Novelty, Recommender Systems

Page ii

Abstract

The explosive growth of the World-Wide-Web and the emergence of ecommerce

are the major two factors that have led to the development of recommender systems

(Resnick and Varian, 1997). The main task of recommender systems is to learn from

users and recommend items (e.g. information, products or books) that match the users’

personal preferences.

Recommender systems have been an active research area for more than a decade.

Many different techniques and systems with distinct strengths have been developed to

generate better quality recommendations. One of the main factors that affect

recommenders’ recommendation quality is the amount of information resources that are

available to the recommenders. The main feature of the recommender systems is their

ability to make personalised recommendations for different individuals. However, for

many ecommerce sites, it is difficult for them to obtain sufficient knowledge about their

users. Hence, the recommendations they provided to their users are often poor and not

personalised. This information insufficiency problem is commonly referred to as the

cold-start problem.

Most existing research on recommender systems focus on developing techniques

to better utilise the available information resources to achieve better recommendation

quality. However, while the amount of available data and information remains

insufficient, these techniques can only provide limited improvements to the overall

recommendation quality.

In this thesis, a novel and intuitive approach towards improving recommendation

quality and alleviating the cold-start problem is attempted. This approach is enriching the

Page iii

information resources. It can be easily observed that when there is sufficient information

and knowledge base to support recommendation making, even the simplest

recommender systems can outperform the sophisticated ones with limited information

resources. Two possible strategies are suggested in this thesis to achieve the proposed

information enrichment for recommenders:

The first strategy suggests that information resources can be enriched by

considering other information or data facets. Specifically, a taxonomy-based

recommender, Hybrid Taxonomy Recommender (HTR), is presented in this

thesis. HTR exploits the relationship between users’ taxonomic preferences

and item preferences from the combination of the widely available product

taxonomic information and the existing user rating data, and it then utilises

this taxonomic preference to item preference relation to generate high

quality recommendations.

The second strategy suggests that information resources can be enriched

simply by obtaining information resources from other parties. In this thesis,

a distributed recommender framework, Ecommerce-oriented Distributed

Recommender System (EDRS), is proposed. The proposed EDRS allows

multiple recommenders from different parties (i.e. organisations or

ecommerce sites) to share recommendations and information resources with

each other in order to improve their recommendation quality.

Based on the results obtained from the experiments conducted in this thesis, the

proposed systems and techniques have achieved great improvement in both making

quality recommendations and alleviating the cold-start problem.

Page iv

Acknowledgements

Thanks to God for giving me such a great opportunity to conduct my PhD

research, and the past four year’s time in my career as a research student was truly joyful

and unforgettable. If I have ever achieved anything in my life, they are not from me but

from God.

I am indebted to a great number of people who kindly offered advice,

encouragement, inspiration and friendship through my time at QUT. Firstly, I would like

to express my utmost gratitude to my principal supervisor and mentor Dr. Yue Xu for

her guidance, her support, for the opportunities she has provided me and for the

invaluable insight she offered me. I am also thankful to my associate supervisors, Dr.

Yuefeng Li and Dr. Richi Nayak, they have provided instrumental inputs and guidance

for my research.

In countless ways, I have received support and love from my family. I would like

to take this opportunity to thank them for all the love, encouragement and wonderful

moments they shared with me over the years. To my mum, for her endless love and

caring, to whom I hope I have given back a fraction of what I have received. To my

brother, Samuel, for providing me support and entertainment. To my father, for his

accompany in my childhood. Finally, I would like to thank my friends and church family,

for all of their supports, prayers and encouragement during my life in Australia.

Page v

Table of Contents

Keywords ................................................................................................................................................. i 

Abstract .................................................................................................................................................. ii 

Acknowledgements ................................................................................................................................ iv 

Acknowledgements ................................................................................................................................ iv 

Table of Contents .................................................................................................................................... v 

List of Figures ...................................................................................................................................... vii 

List of Tables ......................................................................................................................................... ix 

Statement of Original Authorship ........................................................................................................... x 

1 INTRODUCTION ............................................................................................................................. 1 

1.1  Problem Statement ....................................................................................................................... 5 

1.2  Contributions ............................................................................................................................... 6 

1.3  Research Methodology ................................................................................................................ 8 

1.4  Thesis Outline .............................................................................................................................. 9 

2 LITERATURE REVIEW ............................................................................................................... 13 

2.1  Recommender Systems .............................................................................................................. 13 2.1.1  Content-Based Filtering .................................................................................................. 13 2.1.2  Collaborative Filtering .................................................................................................... 16 2.1.2.1 Item-based Collaborative Filtering ................................................................................. 20 2.1.3  Demographic Filtering .................................................................................................... 21 2.1.4  Hybrid Techniques ......................................................................................................... 22 

2.2  Taxonomy-based recommender systems ................................................................................... 26 

2.3  Distributed recommender systems ............................................................................................. 27 

2.4  Evaluating Recommender Systems ............................................................................................ 33 2.4.1  Accuracy Metrics ............................................................................................................ 36 2.4.1.1 Predictive Accuracy Metrics ........................................................................................... 36 2.4.1.2 Classification Accuracy Metrics ..................................................................................... 38 2.4.2  Beyond Accuracy ........................................................................................................... 40 

2.5  Implications ............................................................................................................................... 42 

3 MAKING RECOMMENDATIONS WITH ITEM TAXONOMY ............................................. 45 

3.1  Related work .............................................................................................................................. 48 

3.2  Proposed approach ..................................................................................................................... 49 3.2.1  Notation .......................................................................................................................... 50 3.2.2  Item Preferences based User Clusters ............................................................................. 55 3.2.3  Item Preferences - Taxonomic Preference Relation ....................................................... 58 3.2.4  Extraction of User’s Taxonomic Preferences ................................................................. 59 3.2.4.1 Personal Taxonomic Preference ..................................................................................... 59 3.2.4.2 Cluster Taxonomic Preference ........................................................................................ 66 3.2.4.3 Merge Personal and Cluster Taxonomic Preferences ..................................................... 68 3.2.5  Hybrid Taxonomy Recommender .................................................................................. 69 3.2.6  Cold-Start Proof Hybrid Taxonomy Recommender ....................................................... 75 

3.3  Experiments and evaluation ....................................................................................................... 81 3.3.1  Data Acquisition ............................................................................................................. 82 3.3.2  Verification for Item Preferences - Taxonomic Preference Relation .............................. 82 3.3.3  System Evaluations ......................................................................................................... 86 

Page vi

3.3.3.1 Experiment Framework .................................................................................................. 86 3.3.3.2 Parameterisation ............................................................................................................. 89 3.3.3.3 Evaluation Metrics .......................................................................................................... 91 3.3.3.4 Experimental Results ...................................................................................................... 93 

3.4  Chapter Summary .................................................................................................................... 105 

4 DISTRIBUTED RECOMMENDATION MAKING ................................................................... 107 

4.1  Related work ............................................................................................................................ 108 

4.2  ECommerce-oriented Distributed Recommender .................................................................... 111 4.2.1  General Interaction Protocol ......................................................................................... 119 

4.3  Peer Profiling and Selection .................................................................................................... 125 4.3.1  System Formalisation for EDRS .................................................................................. 126 4.3.2  User Clustering ............................................................................................................. 127 4.3.3  Recommender Peer Profiling ........................................................................................ 128 4.3.4  Recommender Peer Selection ....................................................................................... 132 4.3.4.1 Gittins Indices ............................................................................................................... 132 4.3.4.2 Selection Strategy for EDRS ........................................................................................ 137 4.3.4.3 An Example .................................................................................................................. 138 

4.4  Recommendation Merge .......................................................................................................... 140 

4.5  Experiments and Evaluation .................................................................................................... 144 4.5.1  Data Acquisition ........................................................................................................... 145 4.5.2  Experiment Setup ......................................................................................................... 146 4.5.2.1 Constructing the Recommender Peers .......................................................................... 146 4.5.2.2 Evaluation Metrics ........................................................................................................ 151 4.5.2.3 Benchmarks for the Peer Profiling and Selection Strategy ........................................... 152 4.5.2.4 Simulating the User Feedbacks .................................................................................... 154 4.5.3  Experimental Results .................................................................................................... 155 

4.6  Chapter Summary .................................................................................................................... 159 

5 CONCLUSIONS ............................................................................................................................ 160 

5.1  Contributions ........................................................................................................................... 161 

5.2  Future work .............................................................................................................................. 163 

APPENDIX A: STATISTICAL ATTRIBUTE DISTANCE ......................................................... 165 

APPENDIX B: HYBRID PARITITIONAL CLUSTERING ........................................................ 178 

APPENDIX C: RELATIVE DISTANCE FILTERING ................................................................ 207 

BIBLIOGRAPHY ............................................................................................................................. 223 

Page vii

List of Figures

Figure 1.1. The proposed research method for this thesis. ...................................................................... 8 

Figure 3.1: An example fragment of item taxonomy extracted from Amazon.com. ............................. 54 

Figure 3.2: An example list of items with their taxonomic descriptors. ................................................ 55 

Figure 3.3: Reduce neighbourhood searching space with clustering .................................................... 56 

Figure 3.4. The impact of different values on 2 ( 0.28) ........................................................... 73 

Figure 3.5. The impact of different values on 2 .............................................................................. 73 

Figure 3.6. Recommender evaluation with precision metric ................................................................. 95 

Figure 3.7. Recommender evaluation with recall metric ....................................................................... 96 

Figure 3.8. Recommender evaluation with F1 metric ........................................................................... 96 

Figure 3.9. Computation efficiency results for different recommenders (average seconds per recommendation) .......................................................................................................................... 97 

Figure 3.10. F1 results for HTR with different 1 and configurations. ........................................... 100 

Figure 3.11. F1 results for HTR with different configurations ( 1 0.2) ...................................... 101 

Figure 3.12. F1 results for HTR with different 1 configurations ( 0.8) ...................................... 101 

Figure 3.13. Recommender evaluation under cold-start situations with precision metrics ................. 104 

Figure 3.14. Recommender evaluation under cold-start situations with recall metrics ....................... 104 

Figure 3.15. Recommender evaluation under cold-start situations with F1 metrics ........................... 105 

Figure 3.16. Computation efficiencies for CSHTR and TPR .............................................................. 105 

Figure 4.1. Classical centralised recommender system ....................................................................... 114 

Figure 4.2. Standard distributed recommender system ....................................................................... 116 

Figure 4.3. Proposed distributed recommender system ....................................................................... 119 

Figure 4.4. High level interaction overview for EDRS (based on contract net protocol) .................... 121 

Figure 4.5. The relation between  and Gittins Indices when 0.9 ............................................... 135 

Figure 4.6. Precision results for different recommendation settings ................................................... 158 

Figure 4.7. Recall results for different recommendation settings ........................................................ 158 

Figure 4.8. F1 results for different recommendation settings .............................................................. 159 

Figure A.1. A graph for demonstrating the concept of the standard similarity measures ................... 171 

Figure A.2. A graph for demonstrating the concept of the proposed SAD technique ......................... 172 

Figure A.3. Comparison between IUF and SAD with training sets of different sizes ......................... 177 

Figure B.1. The three major consecutive phases of the proposed HPC technique .............................. 182 

Figure B.2. A possible dataset with a single cluster ............................................................................ 192 

Figure B.3. An example of centroid estimation based on Equation (B.10) ......................................... 192 

Figure B.4. A possible dataset containing multiple clusters ................................................................ 193 

Figure B.5. Centroids estimation for the complex dataset with multiple clusters based on Equation (B.10).......................................................................................................................................... 194 

Page viii

Figure B.6. An example of virtual boundaries for each of the clusters in the dataset ......................... 194 

Figure B.7. An example of cluster centroids estimation process ........................................................ 197 

Figure B.8. Partition quality comparison with different k-means settings .......................................... 200 

Figure B.9. Computation time comparison with different k-means settings ....................................... 201 

Figure B.10. Intra-cluster similarity of the resulting cluster partitions ............................................... 206 

Figure B.11. Inter-cluster distance of the resulting cluster partitions ................................................. 206 

Figure B.12. overall quality of the resulting cluster partitions ............................................................ 206 

Figure C.1. A simple example of the suggested geometrical implication ........................................... 210 

Figure C.2. An example of projected user set ..................................................................................... 211 

Figure C.3. Estimated searching space with three reference users ...................................................... 213 

Figure C.4. An example structure of the RDF searching cache .......................................................... 216 

Figure C.5. Precision Results for different TPR versions ................................................................... 221 

Figure C.6. Recall Results for different TPR versions ........................................................................ 221 

Figure C.7. Average recommendation time for different TPR versions .............................................. 222 

Page ix

List of Tables

Table 3.1. The effect of user clustering on taxonomy information gain ............................................... 86 

Table 3.2. Information for the two different testing datasets ................................................................ 93 

Table 4.1. High level aspect differences among recommender system paradigms ............................. 118 

Table 4.2. The Gittins indices table for 0.9 ................................................................................. 136 

Table 4.3. Performance histories for four recommender peers ........................................................... 139 

Table 4.4. Allocation details for the training and testing user sets ...................................................... 149 

Table 4.5. Dataset allocation details for the four recommender peers ................................................ 150 

Page x

Statement of Original Authorship

The work contained in this thesis has not been previously submitted to meet

requirements for an award at this or any other higher education institution. To the best of

my knowledge and belief, the thesis contains no material previously published or written

by another person except where due reference is made.

Signature: _________________________

Date: _________________________

Page 1

Chapter 1 

1Introduction

The receipt of undesirable or non-relevant information is generally referred to as

information overload (Schafer et al., 2000, Yang et al., 2003). Nowadays, due to the

advancement of internet technology and the World Wide Web (WWW), the issue of

information overload has become increasingly serious. Significant efforts in research are

being invested in building support tools that ensure the right information is delivered to

the right people at the right time. Recommender systems are one of the recent inventions

aiming to help humans deal with this information explosion by giving information

recommendations according to their personal information needs (Linden et al., 2003,

Sarwar et al., 2000b, Schafer et al., 2000). Recommender systems have been applied to

many application areas, including the domain of ecommerce, in which a recommender

system is used to suggest products to customers, and these product suggestions are often

tailored to individual customers’ interests (Linden et al., 2003). Recommender systems

stand out from other information filtering applications in their ability to provide

personalised information recommendations. For example, while standard search engines

are very likely to generate identical search results for users with identical search queries,

recommender systems are able to generate recommendations that are personalised based

on different users’ personal interests (or past behaviours, etc.) even if the users have

identical search queries.

In order to generate personalised recommendations, recommender systems need

to have users’ personal data available. Such personal data includes user demographic

information, user browsing histories, shopping histories, item ratings and user comments.

Page 2

Unfortunately, users’ personal data is difficult to obtain, especially when that data

directly reveal users’ personal interests (e.g. users’ explicit item ratings or comments)

(Park et al., 2006, Schein et al., 2002). Specifically, the situation where a recommender

system has insufficient information resources (e.g. users’ personal data) to generate

quality recommendations is commonly referred to as the cold-start problem (Schein et al.,

2002, Park et al., 2006).

While many of the real world recommender systems suffer from having

insufficient personal data to generate quality personalised recommendations, many

recommender related studies strive to exploit new strategies to better utilise the limited

amount of personal data and information resources to produce better recommendations

(Adomavicius et al., 2005, Badrul et al., 2001, Basu et al., 1998, Deshpande and Karypis,

2004, Goldberg et al., 1992, Jerome and Derek, 2004, Jun et al., 2006). The following

are the main existing strategies for tackling the cold-start problem and improving

recommendation quality:

Developing more sophisticated algorithms to achieve better utilisation of the

limited available information resources (Breese et al., 1998, Montaner et al.,

2003). For example, there are many techniques from other research domains

being applied to recommender systems, such as Bayesian network (Breese et

al., 1998), Neural network (Schafer et al., 2000), and Support Vector

Machine (SVM) (Min and Han, 2005). While some of these advanced

techniques were reported to have achieved better performance, given limited

information resource, the amount of the improvements achieved is often

limited as well.

Hybridising with other techniques that are less dependent on user personal

data (Balabanović and Shoham, 1997, Basu et al., 1998, Burke, 2002). For

Page 3

example, recommenders that are based on users’ personal data can be

combined with standard information filtering techniques, hence, whenever

the recommenders have insufficient personal data to make recommendations

they can use the information filtering techniques as complements to make

recommendations in the case of a cold-start situation. However, such

strategy often risks producing less personalised recommendations.

Even though efforts have been made to improve recommendation quality and

alleviate cold-start problem, no satisfactory solutions have been found so far and the

cold-start problem is still a challenging research problem. This thesis attempts to explore

new strategies to tackle the recommendation making problem – improving

recommendations through information enrichment. As stated earlier, most studies on

recommender systems have been focused on better utilising existing available

information resources, however, very few studies realise that it is also desirable to

increase effectively the amount of information resources available for making

recommendations. In this research, the importance of information enrichment for

recommender systems is highlighted. The objective of this research is to develop

effective strategies to achieve the information enrichment for the recommenders, and

then demonstrate that recommender systems’ performance can be effectively improved

when the available information resources are enriched. Concretely, two novel

recommendation strategies based on the notion of information enrichment are proposed

in this thesis. They are Hybrid Taxonomy Recommender (HTR) and Ecommerce-

oriented Distributed Recommender System (EDRS).

The HTR utilises item taxonomy information with user rating data to make

quality recommendations. One of its major contributions is that it demonstrated the

possibility of integrating user unrelated data (e.g. item taxonomy, item contents, etc.) and

Page 4

users’ personal data (e.g. users’ item ratings and comments) into a useful knowledgebase

that represents users’ interests at a deeper depth. Specifically, HTR extracts the

relationship between users’ item interests and taxonomy interests from the given item

taxonomy information and user rating data, and utilises this relationship to make quality

recommendations. It is shown in our experiment that HTR is able to generate high

quality personal recommendations even under severe cold-start situations. To the best of

our knowledge, there is no similar research that explores the relationship between users’

item preferences and item taxonomic preferences, and exploits this relationship to

produce better recommendations.

The EDRS is a distributed framework that allows multiple recommenders from

different parties (i.e. organisations and ecommerce sites) to cooperate with each other as

well as sharing their information resources and recommendations. While many existing

research on recommender systems focuses on exploring new techniques to better utilise

available information resources, this thesis suggests that if the available information

resources can be enriched, recommenders’ recommendation quality would also be

improved. The cold-start problem would, therefore, be alleviated as well. The idea

behind the proposed EDRS is that instead of improving a recommender’s underlying

algorithm to make better recommendations, the recommender can cooperate with

recommenders from other parties to obtain additional information resources and

recommendations to enrich its available information resources and improve its

recommendation quality. In order to allow the recommenders within the proposed EDRS

to effectively cooperate and interact with each other, a novel recommender peer profiling

and selection strategy is also presented in this thesis. It allows recommenders to learn

from each other and select the most appropriate recommenders to assist in making

recommendations. It is shown in our experiment that by allowing recommenders to

Page 5

cooperate and share their recommendations, their recommendation quality can be

drastically improved. To the best of our knowledge, there is no concept similar to the

proposed EDRS framework in any other research.

Besides the above-mentioned two major contributions (i.e. HTR and EDRS),

three new recommender-related techniques are also developed during this thesis, and

they are Statistical Attribute Distance (SAD), Hybrid Partitional Clustering (HPC), and

Relative Distance Filtering (RDF). These three additional contributions are generic level

techniques designed for improving common recommenders’ recommendation accuracy

and efficiency, and they have been utilised in the development of this thesis. However,

because these three additional studies are not strongly related to the overall theme of this

thesis (i.e. information enrichment), they are not included in the main body of the thesis.

Instead, they are appended as the appendices of this thesis.

To summarise, while many existing studies on recommender systems are about

exploring new techniques to better utilise available information resources, the main

objective of this thesis is to exploit new data resources (i.e. item taxonomy data) and new

system structure (i.e., distributed framework) to achieve information enrichment for

improving recommendation quality and coping with the cold-start problem.

1.1 PROBLEM STATEMENT

Most research on the recommender system community has focused on

developing algorithms to improve recommenders’ recommendation quality, especially in

the situation where only limited information resources are available (i.e. to cope with the

cold-start problem). Majority of the recommender related studies focus on developing

approaches to better utilise the limited available information resources to form better

recommendations. However, given insufficient information resources, the amount of

Page 6

improvements that can be gained from these techniques is very limited. Hence,

improving recommendation quality and alleviating the cold-start problem are still

unresolved problems.

While it is difficult to produce quality recommendations with limited information

resources, it can be easily observed that, if the information resources can be enriched, the

recommendation quality can be drastically improved. The main research problem

involved in this thesis is to explore and develop strategies to achieve the information

enrichment in order to improve recommendation quality and tackle the cold-start

problem.

1.2 CONTRIBUTIONS

This thesis proposes to improve recommenders’ recommendation quality and

tackle the cold-start problem by enriching recommenders’ available information

resources. Two systems are proposed in this thesis, and each of them uses a different

strategy to achieve information enrichment for improving recommendation quality. The

first system, Hybrid Taxonomy Recommender (HTR), utilises the commonly available

product taxonomy information in conjunction with users’ rating data to make quality

recommendations, and it features in strong resistance to cold-start problems. The second

system, Ecommerce-oriented Distributed Recommender System (EDRS), allows

recommenders from different parties to share their information resources and

recommendations with each other and make recommendations cooperatively. EDRS

allows recommenders with insufficient information resources to gain drastic

improvements in their recommendations by providing them with help from other

recommenders. The summarised contributions are briefed as follows:

Page 7

A novel recommender system, HTR, is proposed. It utilises the new

information resource (i.e. product taxonomy information) for making quality

recommendations.

A novel distributed recommender system framework, EDRS, is proposed. It

allows recommender from different parties to share their information

resources and recommendations in a distributed fashion.

A novel recommender peer profiling and selection strategy is proposed to

allow recommenders to learn from each other and achieve more efficient

and effective interactions within EDRS. Overall, by adopting the proposed

peer profiling and selection strategy, the performance of the proposed EDRS

can be effectively improved.

Experimental evaluations are made and the results prove the feasibility and

effectiveness of the proposed HTR and EDRS. Moreover, the experimental

results obtained also suggest that the notion of information enrichment in

recommender systems is significant.

An advanced similarity measure, Statistical Attribute Distance (SAD),

which allows recommenders to more objectively compute the similarities

among user profiles.

A novel clustering method, Hybrid Partitional Clustering (HPC), is proposed.

It allows recommenders to generate efficiently and effectively user or item

clusters. HPC features in its simplicity to use and the ability to update the

clustering results incrementally in accordance to the dataset changes.

A novel neighbourhood formation technique, Relative Distance Filtering

(RDF), is proposed. It allows recommenders to locate efficiently a target

user’s neighbourhood from a large dataset. RDF features in its accuracy,

Page 8

computation efficiency and memory compactness in comparison to other

existing neighbourhood formation techniques.

1.3 RESEARCH METHODOLOGY

Various research approaches have been used in the recommender system field,

and some of these methods include survey, case studies, prototyping and experimenting

(Sarwar et al., 2000a, Herlocker et al., 2004, Schafer et al., 2000). As the research is

considered to focus on the development of new systems or techniques in the

recommender system, and the soundness of these systems, techniques or proposed

strategies have to be supported by the results from the experimentations and evaluations.

Hence, the experimenting approach integrated with the standard information system

research cycle is chosen as the proposed research method. The process of the research

approach used in this research is illustrated in Figure 1.1.

Figure 1.1. The proposed research method for this thesis.

Page 9

1.4 THESIS OUTLINE

The rest of this thesis is summarised as follows:

Chapter 2: This chapter is a literature review of related recommender

techniques, including both conventional and state of the art recommender

systems. In particular, existing studies on taxonomy-based recommenders

and distributed recommenders are reviewed in depth. It pinpoints the current

research on recommender systems and identifies the gap between the

existing recommender studies.

Chapter 3: This chapter presents the proposed Hybrid Taxonomy

Recommender (HTR) and the techniques involved for constructing

knowledgebase from both the taxonomy information and user rating data.

The experimental process involved for evaluating the system and the

experimental results obtained are detailed in this chapter. The relevant

publications about this chapter are:

o Weng, L.T., Xu, Y., Li, Y., and Nayak, R., ‘Improve Recommendation

Quality with Item Taxonomic Information’, Lecture Notes in Business

Information Processing, 2008.

o Weng, L.T., Xu, Y., Li, Y., and Nayak, R., ‘Web Information

Recommendation Making based on Item Taxonomy’, proceedings

of 10th International Conference on Enterprise Information Systems

(ICEIS2008), 20-28, Barcelona, Spain, June. 2008.

(This publication received the best paper award from the

ICEIS2008).

o Weng, L.T., Xu, Y., Li, Y., and Nayak, R., ‘Exploiting Item Taxonomy

for Solving Cold-start Problem in Recommendation Making’, 20th IEEE

Page 10

International Conference on Tools with Artificial Intelligence

(ICTAI2008) , Dayton, Ohio, USA, Nov. 2008.

o Weng, L.T., Xu, Y., Li, Y., and Nayak, R., ‘Improving Recommendation

Novelty Based on Topic Taxonomy’, proceedings of Workshop on Web

Personalization and Recommender Systems (WPRS2007), conjunction

with the 2007 IEEE/WIC/ACM International Conferences on Web

Intelligence and Intelligent Agent, 115-118, Silicon Valley, USA, Nov.

2007.

Chapter 4: This chapter presents the proposed Ecommerce-oriented

Distributed Recommender System (EDRS) and a novel recommender peer

profiling and selection technique that is designed for facilitating the overall

performance of the proposed EDRS. The experimental process involved for

evaluating the system and the experimental results obtained are detailed in

this chapter. The relevant publications about this chapter are:

o Weng, L.T., Xu, Y., Li, Y., and Nayak, R., ‘Distributed Recommender

Profiling and Selection with Gittins Indices’, proceedings

of IEEE/WIC/ACM International Conference on Web Intelligence

(WI2006), 290-293, Hong Kong, China. 2006.

o Weng, L.T., Xu, Y., Li, Y., and Nayak, R., ‘A Fair Peer Selection

Algorithm for an Ecommerce-Oriented Distributed Recommender

System’, accepted by the 4th International Conference on Active Media

Technology, 31-37, Brisbane, Australia, 2006.

o Weng, L.T., Xu, Y., Li, Y., ‘Framework for Ecommerce Oriented

Recommendation Systems’, proceedings of the 4th International

Page 11

Conference on Active Media Technology (AMT05), 19-21 May, 2005,

Japan.

Chapter 5: This chapter concludes this thesis and draws the direction for

future work.

Appendices: In the appendices of this thesis, three novel neighbourhood

formation related techniques designed for helping recommenders to achieve

better recommendation quality and computation efficiency are included. The

relevant publications are:

o Weng, L.T., Xu, Y., Li, Y., and Nayak, R., ‘An Efficient Neighbourhood

Estimation Technique for Making Recommendations’, Lecture Notes in

Business Information Processing, 2008. (Accepted)

o Weng, L.T., Xu, Y., Li, Y., and Nayak, R., ‘Efficient Neighbourhood

Estimation for Recommendation Making’, Proceedings of 10th

International Conference on Enterprise Information Systems

(ICEIS2008), 12-19, Barcelona, Spain, June. 2008.

o Weng, L.T., Xu, Y., Li, Y., and Nayak, R., ‘Efficient Neighbourhood

Estimation for Recommenders with large Datasets’, Proceedings of the

12th Australian Document Computing Symposium (ADCS2007), 92-95,

Melbourne, Australia, Dec. 2007.

o Weng, L.T., Xu, Y., Li, Y., and Nayak, R., ‘A Novel Cluster Centre

Estimation Algorithm with Hybrid Partitional Clustering’, Proceedings

of Data mining International conference (DMIN’07) in the 2007 World

Congress in Computer Science, Computer Engineering, and Applied

Computing (WORLDCOMP’07), June, Las Vegas, USA, 2007.

Page 12

o Weng, L.T., Xu, Y., Li, Y., and Nayak, R., ‘An Improvement to

Collaborative Filtering for Recommender Systems’, Proceedings of the

International Conference on Computational Intelligence for Modelling,

Control and Automation and International Conference on Intelligent

Agents, Web Technologies and Internet Commerce Vol-1 (CIMCA/

IAWTIC2006) , 792-795, Vienna, Austria, Nov. 2005.

o Xu, Y., and Weng, L.T., ‘Improvement of Web Data Clustering Using

Web Page Contents’, Proceedings of the IFIP International Conference

on Intelligent Information Processing (IIP2004), 21-23, Oct., 2004,

Beijing, China.

Page 13

Chapter 2 

2Literature review

This chapter is organised into five sections. Section 2.1 reviews the state of the

art in conventional recommender systems. Section 2.2 summarises recent studies on

recommender systems that exploit the use of item taxonomy or ontology for making

recommendations. Section 2.3 outlines existing studies on distributed recommender

systems. In Section 2.4, various metrics for evaluating the performance of recommender

systems are reviewed. Section 2.5 highlights the implications from the literature

affecting this study.

2.1 RECOMMENDER SYSTEMS

Recommender systems have been an active research area for more than a decade,

and many different techniques and systems with distinct strengths have been developed.

Based on the information filtering (Montaner et al., 2003) techniques employed,

recommender systems can be broadly divided into four categories: content-based

filtering, collaborative filtering, demographic filtering and hybrid techniques. Each of

these categories will be discussed in turn in this section.

2.1.1 Content-Based Filtering

Conventional techniques dealing with information overload typically make use

of content-based filtering techniques. Content-based filtering, also called cognitive

filtering (Malone et al., 1987), relies on charactering the content of an item and

Page 14

information needs of potential users and then using these representations to intelligently

match items to users. In other words, content-based filtering techniques recommend

items with similar contents to the items preferred by target users (Jian et al., 2005,

Pazzani and Billsus, 2007, Malone et al., 1987).

Typically, content-based filtering techniques match items to users through

classifier-based approaches or nearest-neighbour methods.

In classifier-based approaches, each user is associated with a classifier as a

profile. The classifier takes an item as its input and then concludes whether the item is

preferred by the associated user based on the item contents (Pazzani and Billsus, 2007).

Several classifier techniques have been employed in content-based filtering

recommenders, and some of the most common ones are: neural network, decision tree,

rule induction, and Bayesian network. For example, Re:Agent (Boone, 1998) and a

personal news recommender proposed by Jennings (Jennings and Higuchi, 1993) are

based on neural network; Syskill & Webert (Pazzani et al., 1996) and Kim’s

advertisement personalisation technique (Kim et al., 2001) are based on decision tree;

RIPPER (Cohen, 1995, Cohen, 1996), MovieLens (Good et al., 1999), Recommender

(Basu et al., 1998) and WebSIFT (Cooley et al., 1999) are based on rule induction;

News Dude (Billsus and Pazzani., 1999), Personal WebWatcher (Mladenic, 1996) and

Sollenborn’s category-based filtering technique (Sollenborn and Funk, 2002) are based

on Bayesian network.

By contrast, content-based filtering techniques based on nearest-neighbour

methods store all items a user has rated (i.e. expressed his or her interests in the items) in

his or her user profile. In order to determine the user’s interests in an unseen item, one or

more items in the user profile with contents are closest to the unseen item are allocated,

and based on the user’s preferences to these discovered neighbour items the user’s

Page 15

preference to the unseen item can be induced (Montaner et al., 2003, Pazzani and Billsus,

2007). Some of the most well known content-based filtering recommenders utilising

nearest-neighbour methods are: WEBSELL (Cunningham et al., 2001), Daily Learner

(Billsus et al., 2000), LaboUr (Schwab et al., 2000), etc. Content-based filtering

techniques general have the following strengths:

Allow users to get insight into the motivation why the suggested items are

interesting for them since the content of each item is known from its

representation (Montaner et al., 2003).

Content-based filtering techniques are less affected by the cold-start problem

which is one of the major weaknesses of the collaborative filtering based

recommenders.

Generally speaking, purely content-based filtering recommenders have a number

of weaknesses in recommending good items:

Content-based filtering techniques are based on objective information about

the items (such as the text description of an item or the price of a product)

(Montaner et al., 2003), whereas a user’s selection is usually based on the

subjective information of the items (such as the style, quality or point-of-

view of items) (Goldberg et al., 1992). Hence, content-based filtering

techniques generally do not take the user’s perceived valuation of subjective

item information into account when making recommendations. For example,

these techniques might not be able to discriminate between a badly written

and a well written article if both happen to use similar terms.

Content-based filtering techniques often suffer from the over-specialisation

problem. They have no inherent method for generating serendipitous

suggestions, and, therefore, tend to recommend more of what a user has

Page 16

already seen (Resnick and Varian, 1997, Schafer et al., 2000). However, in

many cases, the user’s interests may be beyond the scope of the previously

seen items. Hence, with purely content-based filtering techniques, many

interesting items can hardly be recommended to the user.

In content-based filtering techniques, items need to be represented in a form

such that their semantic attributes can be easily extracted (e.g. text), or

otherwise their attributes will have to be manually assigned. Hence, for

items, such as sound, photographs, art, video or physical items, their

attributes need to be assigned by hand before they can be used in content-

based filtering techniques. However, in many cases, it is not possible or

practical to manually assign these attributes to the items due to limitation of

resources (Shardanand and Maes, 1995).

With purely content-based filtering recommenders, a user’s own ratings are

the only factor influencing the recommenders’ performances. Hence, their

recommendation quality will not be very precise for users with only a few

ratings (Montaner et al., 2003).

Many content-based filtering techniques represent item content information

as word vectors and maintain no context and semantic relations among the

words, therefore the result recommendations are usually very content centric

and poor in quality (Adomavicius et al., 2005, Burke, 2002, Ferman et al.,

2002, Schafer et al., 2000).

2.1.2 Collaborative Filtering

Collaborative filtering, or social filtering, (Malone et al., 1987, Shardanand and

Maes, 1995) is perhaps the most promising technique in recommender systems. It is

Page 17

most known for its use on popular ecommerce sites such as Amazon.com or

NetFlix.com (Linden et al., 2003, Kriss, 2007). Essentially, a collaborative filtering

based recommenders automates the process of ‘word-of-mouth’ paradigm: it makes

recommendations to a target user by consulting the opinions or preferences of the users

with similar tastes to the target user (Breese et al., 1998, Schafer et al., 2000).

Generally, collaborative filtering based techniques provide three major

advantages over other recommendation techniques (especially content-based filtering):

They usually incorporate subjective information about items (e.g. style,

quality, etc.) into their recommendations. Hence, in many cases,

collaborative filtering based recommenders provide better recommendation

quality than content-based recommenders, as they will be able to

discriminate between a badly written and a well written article if both

happen to use similar terms (Montaner et al., 2003, Goldberg et al., 1992).

Collaborative filtering makes recommendations based on other users’

preferences, whereas content-based filtering solely uses the target user’s

preference information. This, in turn, facilitates serendipitous

recommendations because interesting items from other users can extend the

target user’s scope of interest beyond his or her already seen items (Sarwar

et al., 2000b, Montaner et al., 2003).

Collaborative filtering based recommenders are entirely independent of

representations of the items being recommended, and, therefore, they can

recommend items of almost any types including these items that are hard to

extract semantic attributes automatically (e.g. video and audio files)

(Shardanand and Maes, 1995, Terveen et al., 1997). Hence, collaborative

filtering based recommenders work well for complex items, such as music

Page 18

and movies, where variations in taste are responsible for much of the

variation in preferences (Burke, 2002).

Tapestry and GroupLens are the two most widely recognised collaborative

filtering based recommenders. Tapestry (Goldberg et al., 1992, Resnick and Varian,

1997), the earliest implementation of collaborative filtering based recommenders, makes

recommendations based on the explicit opinions of people from a close-knit community

(e.g. an office workgroup ). GroupLens (Konstan et al., 1997) is another widely

recognised recommender system. It computes the correlation between readers of Usenet

newsgroup by comparing their ratings of news stories. An individual user’s ratings are

used to discover other users with similar ratings, and their ratings are processed to

predict the user’s interest in new stories.

Despite their popularity, collaborative filtering based recommenders usually

suffer from the following problems:

One challenge commonly encountered by collaborative filtering based

recommenders is the cold-start problem. Based on different situations, the

cold-start problem can be characterised into two types, namely ‘new-system

cold-start problem’ and ‘new-user cold-start problem’.

The new-system cold-start problem refers to the circumstance where a new

system has insufficient profiles of users. In this situation, collaborative

filtering based recommenders have no basis upon which to recommend, and

hence perform poorly (Middleton et al., 2002).

In the new-user cold-start problem, recommenders are unable to make

quality recommendations to new target users with no or few rating

information. This problem can still happen for systems with a certain

amount of user profiles (Middleton et al., 2002).

Page 19

When a brand-new item appears in the system there is no way it can be

recommended to a user until more information is obtained through another

user rating it. This situation is commonly referred to as ‘early-rater problem’

(Towle and Quinn, 2000, Cöster et al., 2002).

The coverage of user ratings can be sparse when the number of users is

small relative to the number of items in the system (e.g. in a large online

book store there might be tens or hundreds incoming new books everyday).

In other words, when there are too many items in the system, there might be

many users with no or few common items shared with others. This problem

is commonly referred to as ‘sparsity problem’. The sparsity problem poses a

real computational challenge as collaborative filtering based recommenders

may become harder to find neighbours and harder to recommend items since

too few people have given ratings (Gui-Rong et al., 2005, Montaner et al.,

2003).

Another problem is that for users with distinct tastes from others, there will

be no or few other users who share similar tastes to them, and, therefore,

leading to poor recommendations (Montaner et al., 2003).

Scalability is another major challenge for collaborative filtering based

recommenders. Collaborative filtering based recommenders require data

from a large number of users before being effective as well as requiring a

large amount of data from each user while limiting their recommendations

to the exact items specified by those users. The computation efficiency of

collaborative filtering is basically between and , where

is number of users and is number of items (Papagelis et al., 2005). The

numbers of users and items in ecommerce sites might increase dynamically

Page 20

(most of them are over several million), consequently, the recommenders

will inevitably encounter severe performance and scaling issues (Sarwar et

al., 2000a, Gui-Rong et al., 2005, Sarwar et al., 2002).

2.1.2.1 Item-based Collaborative Filtering

Since conventional collaborative filtering based recommenders usually suffer

from scalability and sparsity problems (as described in Section 2.1.2), some researchers

(Badrul et al., 2001, Deshpande and Karypis, 2004, Linden et al., 2003) suggested a

modified collaborative filtering paradigm to alleviate these problems, and this adapted

approach is commonly referred to as ‘item-based collaborative filtering’.

As described in Section 2.1.2, conventional collaborative filtering technique (or

user-based collaborative filtering) operates based on utilising the preference correlations

among users. Unlike the user-based collaborative filtering techniques, item-based

collaborative filtering techniques look into the set of items the target user has rated and

compute how similar they are to the target items that are to be recommended. While

content-based filtering techniques compute item similarities based on the content

information of items, item-based collaborative filtering techniques determine if two

items are similar by checking if they are commonly rated together with similar ratings

(Deshpande and Karypis, 2004). In addition, Lemire and Maclachlan (2005) proposed a

modified item-based collaborative filtering technique called Slope One, and it mainly

features on its computation efficiency and adaptability on user profile changes (i.e. new

ratings are contributed to the dataset). Instead of utilising strongly correlated items in

recommendation making, the Slope One technique is based on the degree of

dissimilarities among the items (in terms of average user preferences). For example, if

Page 21

most people give higher ratings to Item A over Item B, for target users who like Item B it

is very likely that Item A is also preferred by them.

Item-based collaborative filtering usually offers better resistance to data sparsity

problem than user-based collaborative filtering. It is because in practice there are more

items being rated by common users than users who rate common items (Badrul et al.,

2001). Moreover, because the relationship between items are relatively static (compare

to the relationship between users), item-based collaborative filtering can pre-compute the

item similarities offline (where user-based collaborative filtering usually computes user

similarities online) to improve its computation efficiency. Therefore, item-based

collaborative filtering is less sensitive to scalability problem (Badrul et al., 2001, Jun et

al., 2006, Deshpande and Karypis, 2004, Linden et al., 2003).

2.1.3 Demographic Filtering

Demographic filtering techniques employ descriptions of people (e.g. education,

age, occupation, and gender.) to learn the relationship between a single item and the type

of people who like it (Krulwich, 1997, Rich, 1998). For example, when making a book

recommendation to a user with interest to Australian culture, some demographic

information of the user might need to be considered:

The user’s age, occupation or educational background. Is the user an

elementary school student who just needs some introductory textbooks for

his or her homework, or a university professor who needs sophisticated

literatures for research purposes?

The user’s nationality or cultural background. Is the user able to read

English?

Page 22

LifeStyle Finder (Krulwich, 1997) is an example of purely demographic filtering

based recommenders. LifeStyle Finder divided the population of the United States into

62 demographic clusters based on their lifestyle characteristics, purchasing history and

survey responses. Hence, based on a given user’s demographic information, LifeStyle

Finder can deduce the user’s lifestyle characteristics (by finding which demographic

cluster the user belongs to), and make recommendations to the user.

Generally, demographic filtering based recommenders suffer from two principal

shortcomings:

Demographic filtering based recommenders create user profiles by

classifying users using stereotypical descriptors (Rich, 1998). Thus, they

recommend the same items to users with similar demographic profiles.

However, as every user is different, these recommendations might be too

general and poor in quality (Montaner et al., 2003).

Purely demographic filtering based recommenders do not provide any

individual adaptation to interest changes (Montaner et al., 2003). However,

an individual user’s interests tend to shift over time, so the user profile needs

to adapt to change. By contrast, collaborative filtering and content-based

recommenders are generally adaptable to the changes in users’ preferences;

it is because both of them take users’ preference data as input for making

recommendations.

2.1.4 Hybrid Techniques

From the recommendation techniques described in previous sections, it can be

observed that different techniques have their own strengths and limitations, and none of

them is the single best solution for all users in all situations (Wei et al., 2005). A hybrid

Page 23

recommendation system is composed of two or more diverse recommendation

techniques, and the basic rationales of its forming are to gain better performance with

fewer of the drawbacks of any individual technique, as well as to incorporate various

input dataset in order to produce recommendations with higher accuracy and quality

(Schafer et al., 2000). The Active Web Museum, for instance, combines both

collaborative filtering and content-based filtering to produce recommendations with

appropriate aesthetic quality and content relevancy (Mira and Dong-Sub, 2001).

Burke (Burke, 2002) has proposed a taxonomy classifying hybrid

recommendation approaches into seven categories, and they are ‘weighted’, ‘mixed’,

‘switching’, ‘feature combination’, ‘cascade’, ‘feature argumentation’ and ‘meta-level’.

Brief discussions for each of the categories are given below.

‘Weighted’ is the hybridisation method that computes the score of a

recommended item based on summing up the scores that are given to the

item by several recommendation techniques. For example, Funakoshi and

Ohguro (Funakoshi and Ohguro, 2000) have described a simple hybrid

model that uses both collaborative filtering and content-based filtering to

calculate the user similarities, and the recommendations are generated based

on the sum of these two similarities. The benefits of this type of

hybridisation method include low effort and cost on system implementation

and capability of adjusting hybrid weighting.

A ‘switching’ hybrid uses item related criterion to switch between

recommendation techniques. The DailyLearner system (Billsus et al., 2000)

attempted to solve the cold-start problem by employing the content-based

recommendation method first, and if the result recommendations do not

have enough confidence then a collaborative filtering approach is attempted.

Page 24

Deciding the switch criteria is the complexity of switching hybrids, and it

can be determined based on either domain knowledge of the products or

another level of parameterisation. Nevertheless, the advantage of switching

hybrids is they can be sensitive to the weaknesses of their constituent

recommenders (Burke, 2002).

A ‘mixed’ hybrid gathers recommendations from two or more

recommendation techniques and presents them together. This approach is

suitable to be applied in the system where a large number of

recommendations are required. Basically, mixed hybrid systems are very

easy to implement, because it is not necessary to make a reasonable

integration of several techniques, except certain ranking or ordering for the

recommendations must be made. Additionally, care must to be taken to

avoid conflicts and duplications among these mixed recommendations

(Burke, 2002).

‘Feature augmentation’ and ‘Feature combination’ are very similar in the

sense that one recommendation technique’s output is used as an input of

another technique. However, the difference is that the feature augmentation

hybrid requires a staged process whereas the feature combination hybrid

uses a linear approach. An example of feature augmentation hybrid is

described by Popescul and his colleagues (Popescul et al., 2001). They

proposed a new collaborative filtering approach, in which the item ratings

generated through content-based filtering are also used to produce final

recommendations. Feature combination, conversely, works through treating

collaborative information as simply additional feature data associated with

Page 25

each item and then apply content-based filtering technique over this

augmented dataset (Burke, 2002).

A ‘cascade’ hybrid generates recommendations with better qualities by

using one recommendation technique to refine the outputs of another. For

instance, in some cases the relevancy of resulted recommendations of

collaborative filtering is low, and hence content-based filtering can be

employed to filter out the irrelevant recommendations (Burke, 2002).

Another way that allows two recommendation techniques to be combined is

the ‘meta-level’ hybrid, which uses the model generated by one technique as

the input for another. The main difference between ‘meta-level’ and ‘feature

augmentation’ is that a meta-level hybrid uses entire model as the input,

whereas in feature augmentation the input is learnt model. The benefit of

meta-level hybrid is it can solve sparsity problem by compressing ratings

over many techniques into a single model to ease the compressions across

users (Burke, 2002).

The central idea of hybrid recommendation techniques is that they usually

comprise strengths from various recommendation techniques. However, it also means

they might potentially include the limitations from those techniques. Moreover, hybrid

techniques usually are more resource intensive (in terms of computation efficiency and

memory usages) than stand-alone techniques, as their resource requirements are

accumulated from multiple recommendation techniques. For example, a ‘collaboration

via content’ hybrid (Pazzani, 1999) might need to process both item content information

and user rating data to generate recommendations, therefore requires more CPU circles

and memories than any single content-based filtering or collaborative filtering

techniques.

Page 26

2.2 TAXONOMY-BASED RECOMMENDER SYSTEMS

As described in Section 2.1.1, content-based filtering techniques often suffer

from the over-specialisation problem (or content centric problem) because they usually

exploit item content information in word level. To overcome the over-specialisation

problem, taxonomy-based techniques, therefore, are proposed to use item taxonomic or

semantic information to make information filtering process more meaningful (Hollink et

al., 2007). For example, for a target user interested in ‘flower’, content-based filtering

techniques might only consider items with exact word ‘flower’ in the content, whereas

taxonomy-based techniques might also consider items related to words such as ‘rose’,

‘seeds’, etc..

The application of taxonomic information in information filtering related tasks

has been explored before. The most well-known example is the directory based

browsing of information mines, for example, ACM Computing Reviews

(http://www.reviews.com/), Google Directory (http://directory.google.com/) and Yahoo

(http://www.yahoo.com/). These sites organize their information items (e.g. web pages)

based on the items’ taxonomic information, and allow users to easily locate desired items

by browsing and traversing the taxonomic structure imposed by these items’ taxonomic

information. Moreover, category based filtering techniques have been proposed (Kohrs

and Merialdo, 2000, Sollenborn and Funk, 2002) that put emphasis on categories as

meta-data to improve recommendation qualities as well as computation efficiency.

Pretschner and Gauch (1999) proposed a personalised web search technique with

ontology based user profiling. The CHIP Demonstrator (Aroyo et al., 2007) also makes

semantics-driven recommendations by allowing users to explicitly rate a set of

predefined semantic attributes of the items. The E-Culture Demonstrator alleviates over-

Page 27

specialisation problem by expanding users’ searching queries with word semantics

(Hollink et al., 2007).

There are also some studies that specifically consider utilising item taxonomic or

ontological information to assist recommender systems. Middleton et al (2002) use

ontology to inductively learn user interested topics for recommending research papers to

users. Based on the set of user-interested topics, the recommendation list can be

efficiently generated by weeding out those research papers that do not fall into these

preferred topics. Conversely, Ziegler et al (2004) proposed a taxonomy-driven product

recommender, it utilises a general tree structured product taxonomy to enhance its

recommendations.

Most of the current studies are based on mapping the target user’s taxonomic

(semantic or ontological) interests against other user’s taxonomic interests (for forming

neighbourhoods), or against the taxonomic information of the items (for information

filtering or recommendation making). As such, their underlying logic is similar to

conventional content-based filtering techniques. However, because taxonomic

information is sophisticate and information rich, there are still many potential promising

ways to utilise it in the applications of information filtering and recommender systems.

2.3 DISTRIBUTED RECOMMENDER SYSTEMS

To date, many recommender systems have been crafted with centralised

scenarios in mind; that is, assuming recommenders can access, retrieve, utilise all data

and information (e.g. user browsing/rating histories and product information) from a

centralised database or data repository (Liu et al., 2007). Centralised recommenders have

been popularly applied in Business to Customer (B2C) applications (especially these

ecommerce websites such as Amazon.com, Book.com, etc.), as they generally adhere to

Page 28

client-server architecture where centralised recommenders and data repositories are

hosted by the central server. A detailed review of centralised recommenders is provided

in Section 2.1 and Section 2.2.

Notwithstanding the popularity of centralised recommenders in last decades,

recommender systems that operate on distributed environments or decentralised

infrastructures have begun to attract attention from researchers, and these systems are

commonly referred to as distributed recommender systems or decentralised

recommender systems (Castagnos and Boyer, 2007, Clements et al., 2007, Liu et al.,

2007).

Generally, a distributed recommender system associates each of its users with a

recommender agent (or peer recommender) on his or her personal computer (client-side

machine). These recommender agents gather user profile information from their

associated users, and exchange these profile information with other agents over a

distributed network (e.g. the internet). In the end, a recommender agent makes

recommendations to its associated user by utilising the user’s personal profile as well as

these gathered peer profiles (i.e. profiles of other users gathered from other

recommender agents) (Castagnos and Boyer, 2007, Han et al., 2004, Tveit, 2007, Vidal,

2004, Wang et al., 2006).

There are several reasons that have led the increasing popularity of distributed

recommender systems:

The fast growing development of internet related technologies and

applications (e.g. the Grid, ubiquitous computing, peer-to-peer networks for

file-sharing and collaborative tasks, Semantic Web, social communities,

WEB 2.0, etc.) has yielded a wealth of information and data being

distributed over most of nodes (i.e. web server, personal computer, and

Page 29

mobile phone) in the internet. Hence, getting information recommended

from only one single source (e.g. ecommerce site) is no longer sufficient for

many users, and instead, they are thirsty for richer information from multiple

sources (Han et al., 2004, Miller et al., 2004, Tveit, 2007). For example, the

peer-to-peer (P2P) based file sharing protocol, BitTorrent

(www.bittorrent.com), has proven to be among the most competent methods

to allow large numbers of users to share efficiently large volumes of data.

Instead of storing files or data in a central file server (e.g. FTP server),

BitTorrent stores files in multiple client machines (i.e. peers), and when a

file is requested by a user (i.e. a peer), the user can download this file

simultaneously from multiple peers (Clements et al., 2007). Intuitively, as

there is no central server for storing file contents and user (or peer) profiles

in BitTorrent, distributed recommender systems would be more suitable to

be applied to such system than centralised recommenders.

User privacy and trust is another area that distributed recommender systems

are considered superior to centralised recommender systems. In a centralised

recommender system, all user information and profiles are possessed by the

ecommerce site that runs the recommender system, and this can result in two

privacy and trust concerns. Firstly, a centralised recommender system might

share users’ personal information and profile in inappropriate ways (e.g.

selling user information to others), and the users generally have no control

over it. Secondly, a centralised recommender system owned by an

ecommerce site might make recommendations for the business’s own good

instead of serving users’ needs. For example, a site can adjust its

recommender’s settings, so it only recommends products that are

Page 30

overstocked instead of those required by the users (Foner, 1997, Miller et al.,

2004).

The privacy and trust issues are alleviated by distributed recommender

systems. In a distributed recommender system, users’ personal information

and profiles are stored in their own machines, and they generally can

explicitly define and set which parts of their personal data and profiles are

sharable. In addition, because a recommender agent in a distributed

recommender system is a piece of software that runs independently on each

client’s machine and it usually gather information only from other peer

agents rather than from an ecommerce site, therefore, it is less possible that

ecommerce sites can manipulate recommendations to the users (Miller et al.,

2004).

As mentioned in Section 2.1, scalability is one of the major challenges of

centralised recommender systems, and it is because correlating user interests

in a large dataset can be very computationally expensive (it normally require

a quadratic-order matching steps). Some researchers, therefore, suggested

implementing recommender systems in a decentralised fashion to improve

the scalability and computation efficiency (Foner, 1997, Han et al., 2004,

Tveit, 2007).

Yenta (Foner, 1997), a referral-based matchmaking system for online

communities, is often recognised as the first distributed recommender system. Yenta

learns a user’s interests and represents a user’s profile with a set of keywords, based on

the user profile Yenta then match the user with other people with similar interests (by

comparing the keywords of their user profiles). Strictly speaking, Yenta is not

specifically designed for recommendation making, however, because its central idea

Page 31

‘finding like-minded neighbours in distributed environments’ is strongly related to the

concept of distributed recommender systems, many researchers still consider it as the

foundation of distributed recommender systems (Miller et al., 2004, Ogston et al., 2003,

Sorge, 2007, Wang et al., 2006). Additionally, several recent studies focus on distribute

neighbourhood formation are described in (Clements et al., 2007, Link et al., 2005,

Ogston et al., 2003).

Besides grouping users based on the similarities of users’ interests, in order to

prevent crime and improve security for distributed recommender systems, the concept of

trust has been suggested as another factor to be considered when forming user

neighbourhoods (Sorge, 2007, Han et al., 2004, Miller et al., 2004). Moreover, because

trust model imposes another filtering layer, it is also suggested that the computation

efficiency and scalability of distributed recommender systems can, therefore, be

improved (Ziegler and Golbeck, 2007).

While distributed neighbourhood formation is the major research focus in the

field of distributed recommender systems, there are still many other associated

challenges (e.g. communication protocols, decentralised ranking and profile merging) in

the field waiting for more attention. The first complete architecture and protocol for

distributed recommender systems is proposed by Vidal (2004). Some other references

focusing on system architecture and design of distributed recommender systems can be

found in (Castagnos and Boyer, 2007, Liu et al., 2007, Sorge, 2007, Tveit, 2007, Wang

et al., 2006, Yang et al., 2007).

Despite the growing popularity, distributed recommender systems are generally

considered more complex and sophisticated than centralised recommender systems, as

they usually operate in a distributed environment and involve other research disciplines,

Page 32

such as multi-agent systems, grid computing, and distributed systems. In general,

distributed recommender systems impose the following three research challenges:

Neighbour discovery and selection. As distributed recommender systems

mainly operate in a distributed environment, it is assumed each

recommender peer (or agent) operates autonomously and might not (or

cannot) know about every other agents, peers, users, or resources on the

network. Hence, the task of finding like-minded peers in distributed

recommender systems is much harder than in centralised recommender

systems, as distributed recommender systems are required to consider the

various differences (e.g. user profile domain and representation, and

communication protocol) among these autonomous peers. Moreover,

because communications over a distributed environment (e.g. internet) can

be very expensive and inefficient, therefore, the communication traffics and

efficiencies are also essential factors to be considered when designing

strategies for distributed neighbour discovery and selection (Foner, 1997,

Ogston et al., 2003).

Recommendation accuracy. As mentioned previously, finding like-minded

peers is difficult for distributed recommender systems. It is very common

that the discovered neighbours are not globally optimal, and, therefore,

results poor recommendations. In particular, when a distributed

recommender system is in its initialisation (or bootstrapping) phase, each

recommender peer in the system is randomly assigned with a set of initial

neighbour peers. It takes a reasonable amount of time for each peer to learn

and explore other peers in the system, hence, it is difficult for recommender

peers to achieve satisfactory recommendations in such starting stage (Miller

Page 33

et al., 2004, Ogston et al., 2003, Yang et al., 2007). Additionally, because

the recommender peers operate autonomously, it is not possible to expect

that all peers are accessible at any given point of time. As a recommender

agent’s performance mainly depends on the presences of other agents,

maintaining the stability of the recommendation quality in distributed

recommender systems can be challenging (Castagnos and Boyer, 2007,

Foner, 1997).

User privacy and trust. As mentioned previously, distributed recommender

systems can potentially protect users’ privacy as well as avoid manipulated

recommendations from malicious commercial site owners. However,

distributed recommender systems can still suffer from privacy abuses and

recommendation manipulations among the recommender peers (Sorge, 2007,

Castagnos and Boyer, 2007, Chen et al., 2000, Link et al., 2005). For

example, a malicious user can register and construct multiple recommender

peers in a distributed recommender system, and create multiple fake user

profiles to manipulate recommendations generated to their neighbours.

Moreover, it is also possible for the user to use the collected neighbour

profiles for revealing their real world identities and abuse their privacy

(Sorge, 2007).

2.4 EVALUATING RECOMMENDER SYSTEMS

Recommender systems have been an active research area for more than a decade,

and, therefore, many different techniques and systems have been suggested and

developed. In order to select a recommender system that is most suitable for a given

application domain from amongst all other alternative recommender systems, well

Page 34

defined metrics and measures are required for evaluating and comparing these

recommenders (Herlocker et al., 2004).

In the broadest sense, a recommender system can be evaluated for its

recommendation quality or computation efficiency. In the recommendation quality

evaluation, a recommender is evaluated based on whether its recommendations can

satisfy users’ information needs. In other words, if the recommender’s recommendation

quality is good, then they would make most of its users happy and satisfied (Herlocker et

al., 2004). On the other hand, the computation efficiency evaluation aims for ensuring a

recommender’s ability to handle a large number of recommendation requests in real time

(Rashid et al., 2006b, Rashid et al., 2006a, Sarwar et al., 2000a, Sarwar et al., 2002).

Specifically, a common approach to evaluate a recommender’s computation efficiency is

to measure the amount of time it required to generate a single recommendation. In

general, most studies in this field consider a recommender’s recommendation quality

over its computation efficiency. It is because while recommendation quality can only be

improved algorithmically, the efficiency bottleneck can be solved by other non-

algorithmic approaches (such as employing higher performance hardware) (Karypis,

2001, Sarwar et al., 2000b).

Depending on the types of source information employed to evidence whether a

recommendation is preferred by a given user in the evaluation process, existing

evaluation approaches can be divided in to two categories, namely off-line evaluation

and on-line evaluation (Hayes et al., 2002). In off-line evaluation, the performance of a

recommender system is evaluated on existing datasets. In on-line evaluation,

performance is evaluated on users of a running recommender system (Hayes et al., 2002,

Herlocker et al., 2004). Most existing studies on recommender system employ off-line

evaluation rather than on-line evaluation, and it is because:

Page 35

On-line evaluation requires a fully engineered system with certain amount of

online users available to test the system. However, these two requirements

are cumbersome and difficult to achieve for many research projects (Hayes

et al., 2002, Herlocker et al., 2004).

On-line evaluation requires users to actively provide feedbacks to given

recommendations, however, there is high possibility that users might not

provide feedbacks or even give false feedbacks. In general, most users

refuse to provide feedbacks to recommendations as it does not reward them

immediately (Montaner et al., 2003, Pazzani, 1999).

Off-line evaluation, in contrast to on-line evaluation, has the advantage that

it is economical and quick to conduct large scope evaluations (i.e. running

several dataset, metrics and recommendation algorithms at once) (Herlocker

et al., 2004).

Despite the popularity of off-line evaluation, it still suffers from some drawbacks:

The set of items that can be evaluated in off-line evaluation is limited by the

natural sparsity of ratings in datasets. Given a recommended item that has

not been seen by the target user, it cannot be judged that if the item will be

preferred by the user or not (Herlocker et al., 2004).

Off-line evaluation is limited to objective evaluation of prediction results. In

off-line evaluation, it is not possible to determine whether users will prefer a

particular system, either because of its predictions or because of other less

objective criteria such as the aesthetics of the user interface (Herlocker et al.,

2004).

Due to the limited scope of this thesis, only off-line evaluations are carried out

for all recommender related experiments here. The following sections review some

Page 36

popular off-line evaluation metrics for evaluating the recommendation qualities of

recommenders.

2.4.1 Accuracy Metrics

Most of studies on recommender systems evaluate recommendation quality

through measuring the recommendation accuracy, and these techniques for the accuracy

measurements are commonly referred to as accuracy metrics (Herlocker et al., 2004).

Predictive accuracy metrics and classification accuracy metrics are the two major types

of accuracy metrics. Predictive accuracy metrics generally are used to measure how well

a recommender can predict a user’s exact rating value to a specific item. On the other

hand, classification accuracy metrics measure a recommender’s ability to select high

quality items from the set of all items for a given target user (Herlocker et al., 2004,

Montaner et al., 2003, Ziegler et al., 2004).

2.4.1.1 Predictive Accuracy Metrics

In general, predictive accuracy metrics compute the difference between the

predicted user ratings and the true user ratings for a given set of items. Hence, predictive

accuracy metrics are particularly important for recommenders whose tasks are to display

the rating predictions to users (Herlocker et al., 2004, Montaner et al., 2003). For

example, the recommendation task for Tapestry (Goldberg et al., 1992) and GroupLens

(Resnick and Varian, 1997) is to explicitly provide predicted ratings associated with each

posting in a structured posting forum to indicate target users which postings are worth

reading. Thus, for the evaluation of these two recommenders predictive accuracy metrics

are applied.

Page 37

Mean absolute error (MAE) is perhaps the most prominent and widely used

predictive accuracy metric (Breese et al., 1998, Good et al., 1999, Herlocker et al., 2002,

Herlocker et al., 2004, Shardanand and Maes, 1995), and it is the average differences

between the predicted and actual ratings for a given set of items, specifically:

| |∑ | , , |

| | 

(2.1)

where is a subset of items that a user has rated before. , and , denotes ’s

actual and predicted ratings to item respectively. It can be observed from the

equation if the rating predictions are accurate, the value of MAE (i.e. | |) will be small,

conversely, large value of MAE indicates inaccurate rating predictions.

Additionally, there are several variations to the MAE, such as mean squared

error (MSE), root mean squared error (RMSE), and normalised mean absolute error

(NMAE). MSE and RMSE square the differences between the actual and predicted

ratings before summarising them and hence their results emphasise large prediction

errors. For example, misprediction of 2 points increases the MAE value only by 4, but

misprediction of 3 points increases the MAE value by 9 (i.e. mispredictions in extreme

cases are treated seriously) (Herlocker et al., 2004). Another metric, NMAE, was

discussed by Goldberg et al.(1992), it is mean absolute error normalised with respect to

the range of rating values, and it allows comparison between prediction runs on different

datasets.

Page 38

2.4.1.2 Classification Accuracy Metrics

Classification accuracy metrics measure the frequency with which a

recommender system makes correct or incorrect decision about whether an item is good.

They are also referred to as decision support metrics (Herlocker et al., 2004, Ziegler et al.,

2004). Classification accuracy metrics are commonly used for evaluating recommenders

whose tasks are to recommend a ranked list of the recommended items (i.e. a set of all

good items) (Linden et al., 2003, Shardanand and Maes, 1995, Wei et al., 2005, Ziegler

et al., 2004, Deshpande and Karypis, 2004, Karypis, 2001).

Nowadays, recommenders designed specifically for making item list based

recommendations are very popular, and, therefore, classification accuracy metrics have

been widely applied and many different variations have been developed. Among all

different variations, precision and recall are the most basic classification accuracy

metrics. Precision and recall were initially suggested by Cleverdon in 1966 (Cleverdon

et al., 1966) as evaluation metrics for information retrieval systems. Due to the simplicity

and the popular uses of these two metrics, they have been widely adopted for

recommender system evaluations (Basu et al., 1998, Billsus and Pazzani., 1999, Sarwar

et al., 2000a, Sarwar et al., 2000b, Ziegler et al., 2004). Precision and recall for an item

list recommended to user are computed based on the following equations:

Recall| |

| | 

(2.2)

Page 39

Precision| || |

(2.3)

where is the set of all items preferred by user , and is the set of all recommended

items (generated by the recommenders). Based on the Equation (2.2) and (2.3), it can be

observed that the values of precision and recall are sensitive to the size of the

recommended item list (i.e. | |), that is, when the size of the recommended item list is

large, the result precision will be small and recall will be large. In contrast, when the size

of the recommended item list is small, the result precision will be large and recall will be

small.

Since precision and recall are inversely correlated and are dependent on the size

of the recommend item list, they must be considered together to evaluate completely the

performance of a recommender (Herlocker et al., 2004). F1 metric suggested by Sarwar

et al. (2002) is one of the most popular techniques for combining precision and recall

together in recommender system evaluation, and it can be computed by the following

formula:

F12

(2.4)

While precision, recall and F1 metrics are directly adopted from the field of

information retrieval, many of their variants have been suggested for better applicability

in the context of recommender systems. Breese score (also known as weighted recall) is

Page 40

one such notable example. Breese score is proposed by Breese et al. (1998), and it

considers the fact that items at the end of a recommendation list are less likely to be

viewed by the active user. Hence, the quality (i.e. the obtained Breese score) of a

recommendation list also depends how items are arranged in the list. Other popular

classification accuracy metrics include Relative Operating Characteristic (ROC) and

Customer ROC (CROC) metrics (Herlocker et al., 2002, Schein et al., 2002), these two

metrics measure the extent to which an information filtering system is able to distinguish

between signal (user preferred items) and noise (user unseen or disliked items). In

contrast, the NDPM metric employed by FAB recommender system (Balabanović and

Shoham, 1997) considers predictive accuracy for items in the recommendation lists (i.e.

combines both predictive accuracy and classification accuracy metrics), and, therefore,

imposes a higher standard for recommendation list evaluations.

2.4.2 Beyond Accuracy

Although recommendation accuracy is an important facet for recommender

system evaluation, there are still many other factors that can affect users’ satisfaction and

perceptions to a recommender. For example, a recommender might achieve high

accuracy by only recommending popular items; however, some users might find such

recommenders rather boring and expect for recommendations with serendipity. The

following lists some of other facets for recommender evaluation:

Coverage. A recommender with good coverage indicates it is able to make

predictions on most items. Recommenders with lower coverage may be less

valuable to users, because they will be limited in the decisions they are able

to help with (Herlocker et al., 2004). Coverage measure has been the most

popular metric among all non-accuracy based evaluation metrics, and it

Page 41

measures the percentage of elements part of the problem domains (i.e. items

or item categories) for which predictions can be made (Good et al., 1999,

Herlocker et al., 2004, Middleton et al., 2004).

Novelty and Serendipity. Recommenders with novelty and serendipity are

able to make non-obvious recommendations. Some recommenders produce

highly accurate recommendations (i.e. obtain high scores with accuracy

metrics) may still be useless in practice if their recommendations are too

obvious. For example, a recommender in a grocery store might suggest milk

to any shopper who has not yet selected it. Statistically, this recommender is

highly accurate as almost everyone buys milk when they are grocery

shopping. However, such a recommendation is not very useful, because

everyone who comes to grocery store to shop has bought milk in the past,

and knows whether or not they want to purchase more (Herlocker et al.,

2004).

Novelty and serendipity metrics measure the degree to which the

recommenders are presenting items that are both attractive to users and

surprising to them. However, designing these metrics is difficult because

usual methods for measuring accuracy are directly antithetical to novelty

and serendipity. In fact, even though novelty and serendipity have started

getting attentions from researchers (Schafer et al., 2000, Ziegler et al., 2004),

no standard metric for evaluating novelty and serendipity of recommenders

is yet available.

Learning Rate. Given recommenders with similar recommendation

accuracy, it should be obvious that the one requires least amount of data or

information (e.g. rating data) should be superior to others. In general,

Page 42

learning rate metrics measure the amount of information a recommender is

required to produce recommendations with a certain level of accuracy

(Herlocker et al., 2004). Based on different information types, there are three

different learning rate metrics: overall learning rate, per-item learning rate,

and per-user learning rate. The overall learning rate measures the amount of

overall ratings required by a recommender to produce quality

recommendations. The per-item learning rate measures the amount of

ratings to an item are required to allow accurate rating prediction to the item.

The per-user learning rate measures the amount of ratings from an user are

required to allow quality recommendations generated to the user (Herlocker

et al., 2004).

2.5 IMPLICATIONS

In Section 2.1, several classic and state of the art recommender systems are

reviewed. Based on the review, three major information resources employed by

recommender systems for recommendation making are identified, and they are:

item content information (Section 2.1.1)

user demographic data (Section 2.1.3)

users’ past browsing, shopping and rating histories (Section 2.1.2)

Among the three information resources, user rating data is considered the most

popular as it directly relates to users’ personal preferences. However, user rating data is

sometimes difficult to obtain, especially for these new and small ecommerce sites. The

lack of information resources could subsequently affect recommenders’ performances,

and hence result in the cold-start problem (Section 2.1.2 and 2.1.4). Except for

alleviating the cold-start problem in algorithm level (Section 2.1.4), the more

Page 43

fundamental solution is to enrich the information resources. There are two basic ways to

accomplish the enrichment of information resources:

Enrich the information resources by considering other facets of the data.

Enrich the information resources by obtaining more data.

Section 2.2 reviewed a classic example of how other data facets can be utilised to

improve recommendation quality. Researchers have recently suggested that the cold-

start problem can be effectively alleviated by considering item taxonomic information

into recommendation making process (Aroyo et al., 2007, Ziegler et al., 2004). Item

taxonomic information has started getting attentions due to the increasing popularity of

semantic web and ontology related research, and it is considered more sophisticated,

well structured and widely applicable than standard item content information (e.g.

keywords vectors). However, the application of taxonomic information in recommenders

is still relatively new, and most taxonomy-based recommenders simply treat the item

taxonomic information as ordinary content information. Therefore, we believe there is

still a large gap in the effective utilisation of item taxonomic information, and one of the

major goals of this thesis is to explore other promising utilisation of item taxonomic

information to alleviate cold-start problem as well as improve recommendation quality.

One of the most intuitive ways to increase the data volume for a recommender is

to obtain data from other parties, especially from other recommenders. While

recommenders mainly operate over internet, in order to automate the data gathering

progress, it is required to allow multiple recommenders communicate and exchange data

in a decentralised fashion. Hence, studies related to distributed and decentralised

recommender systems were reviewed and investigate (Section 2.3). Based on the review,

most distributed recommender systems are designed for peer-to-peer based applications

and their goal is to move the ownership of recommenders from site owners’ hand to

Page 44

individual user’s hand (i.e. change from B2C to C2C). However, we have not found any

studies on distributed recommenders that address how the cooperation of multiple

recommenders over distributed network can enhance each other’s recommendation

quality as well as alleviate the cold-start problem. As the goal of this thesis is to

investigate novel techniques for alleviating the cold-start problem, it also investigates the

possibility of distributed information sharing for improving their recommendation

quality and resistance to the cold-start problem.

As several novel recommendation techniques are proposed, investigated and

developed in this thesis, it is important to evaluate them and compare them with other

existing techniques. Therefore, in Section 2.4, state of the art evaluation metrics and

various recommender evaluation aspects are reviewed.

Page 45

Chapter 3 

3Making Recommendations with Item

Taxonomy

As mentioned in Chapter 1, one of the major issues that recommender systems

are facing is the cold-start problem. This problem often arises in the following situations:

The target user has very few ratings (e.g. a new user). In this scenario,

recommenders (especially collaborative filtering based recommenders)

might not be able to find users with tastes that are truly similar to the target

user, thus, the quality of recommendations to the target user might be poor.

Moreover, it is difficult to obtain the content interests of the target user

because of the very limited number of items rated by the target user.

The amount of explicit rating data in the system is small. Many

recommender systems rely on the explicit ratings to find users with similar

item preferences to the target user. In the case of lacking sufficient rating

data, recommenders may not be able to find similar users and hence to make

quality recommendations.

It can be observed that the major cause for the above two situations is the heavy

dependency and reliance on explicit item rating data for recommendation making.

Indeed, most recommender systems (especially collaborative filtering based ones) make

recommendations based on users’ item preferences, and the item preferences are mainly

extracted from these users’ explicit item rating data. When the amount of explicit rating

Page 46

data is insufficient, the induced item preferences for the users may be, therefore,

inaccurate, and this consequently leads to poor recommendation quality.

A user’s item preference reflects the user’s perceptions to the quality of the items

that he or she has seen or observed. Hence, with the proper use of the item preference

information and collaborative filtering techniques, a user’s potential perception to a

given item’s quality can be predicted. However, a user’s satisfaction to a given

recommended item (or a list of recommended items) may not solely depend on whether

the quality of the item matches the user’s true perception to the item (Herlocker et al.,

2004). There are many other factors that may affect the user’s perception to a given

recommendation, such as the size and the order of the items of the recommended item

list, the novelty or serendipity of the recommendation, the taxonomic relevance of the

recommendation to the user’s taxonomic interests, etc (Herlocker et al., 2004). Hence, in

order to maximise the user’s satisfaction, recommenders should utilise other information

resources rather than solely rely on the explicit rating data.

In this chapter, we explore a new information resource – item (or product)

taxonomic information – to alleviate cold-start problems as well as improve

recommendation quality. Item taxonomy is a set of controlled vocabulary terms or topics,

usually hierarchical, designed to describe and classify items (Levy, 2004). Due to the

drastic growth of information volume, ecommerce sites and Business-to-Business (B2B)

applications, the development and application of item taxonomy are becoming

increasingly popular. For example, the United Nations Standard Products and Services

Classification (UNSPSC) specifies more than 11,000 taxonomy codes and the

hierarchical order to describe and classify products and services for use throughout the

global marketplace (Levy, 2004, Leo et al., 2003). Ecommerce sites such as

Amazon.com (http://www.amazon.com), BARNES&NOBLE (http://www.book.com),

Page 47

art.com (http://www.art.com), eBay (http://www.ebay.com) etc. also provide their own

item/product taxonomy to describe and classify their goods.

This thesis exploits item taxonomic information to obtain users’ taxonomic

preferences from their past ratings and browsing histories. A user’s taxonomic

preferences reflect the user’s interests to the category or catalogue of the items. The main

differences between users’ taxonomic preferences and item preferences is that item

preferences capture users’ perceptual tastes to items, whereas taxonomic preferences

capture users’ content interests to items. Instead of only using users’ item preferences

like what the standard collaborative filtering does, we make use of both users’ item

preferences and taxonomic preferences. In the cases of lacking rating data or for a new

target user, even there might be no similar users according to the target user’s item

preferences, we still can find users who have similar taxonomic preferences with the

target user. Moreover, because we are able to obtain users’ taxonomic preferences from

both their explicit and implicit ratings, we can assure there is sufficient user taxonomic

preferences information for generating quality recommendations even when the amount

of user explicit rating data in the system is small.

This chapter presents two recommendation techniques that make use of item

taxonomic information. The first technique, which is called Hybrid Taxonomy

Recommender (HTR), utilises item taxonomic information to improve the

recommendation quality of standard item-based collaborative filtering systems. The

second technique, which is called Cold-Start Proof Hybrid Taxonomy Recommender

(CSHTR), is developed specifically for systems operating in environments with severe

cold-start problems.

Page 48

3.1 RELATED WORK

Much research has suggested that the cold-start problem can be alleviated by

combining collaborative filtering and content-based techniques (Burke, 2002, Ferman et

al., 2002, Park et al., 2006, Schein et al., 2002). However, as part of the recommendation

process for these hybrid recommenders is content-based, the generated recommendations

may be excessively content centric and lack of novelty (Middleton et al., 2002, Ziegler et

al., 2004). Hence, semantic and ontology based techniques have been suggested to

improve the recommendation generality for the content-based filtering. Middleton

(Middleton et al., 2002) suggested an ontology based recommender which uses external

organisational ontology (e.g. publication-and-authorship relationships, and projects-and-

project membership relationships) to solve the cold-start problem. However, as the

Middleton’s technique is mainly designed for recommending research papers and

documents, and relies on a specific organisational ontology, therefore, it is not easy to

adopt this method for general recommenders. Another work is the taxonomy-driven

product recommender (TPR) proposed by Ziegler et al (Ziegler et al., 2004). TPR utilises

a general, tree structured product taxonomy to enhance its recommendations. Due to the

simplicity of the taxonomy structure, Ziegler’s technique is considered widely applicable

to different domains (Ziegler et al., 2004). To the best of our knowledge, Middleton and

Ziegler’s techniques are the only two studies bearing traits similar to the proposed HTR

and CSHTR techniques. HTR and CSHTR employs similar tree structured taxonomy

used in TPR, and, therefore, they inherit TPR’s generality advantage. However, TPR is

only applicable to use user’s implicit taxonomic preferences for making

recommendations, whereas HTR and CSHTR utilise the relationship between users’

implicit taxonomic preferences and explicit item preferences for recommendation

making, therefore, yielded better recommendation performance and work well in the

Page 49

case of lacking of implicit taxonomic data. Moreover, HTR and CSHTR inherit item-

based collaborative filtering paradigm (Deshpande and Karypis, 2004) (in contrast to

TPR’s user-based collaborative filtering), therefore, most computations can be done

offline which results in significant improvement to the computation efficiency of online

recommendation generation.

3.2 PROPOSED APPROACH

The basic idea behind HTR is intuitive. It firstly finds a set of users (i.e. the

neighbours) with similar item preferences to a given target user, and then extracts

taxonomy topics that are popularly and uniquely preferred by these users. By combining

the taxonomy topics preferred by the target user and his/her neighbours, the taxonomic

preferences of the target user is induced. Finally, HTR estimate the target user’s

preference to a candidate item by combining his/her item preferences with taxonomic

preferences. By utilising both the users’ item preferences and item taxonomic

preferences, HTR offers two major advantages over other existing recommenders based

on only item preferences. Firstly, when two items are both preferred by the target user’s

neighbours, HTR will assign higher score to the item whose taxonomy topics are more

popularly and uniquely preferred by the neighbours. Since extra information resources

(i.e. users’ item taxonomic preferences) are utilised to refine the recommendations,

therefore, better recommendation quality is achieved in HTR. Secondly, for items with

only few or no ratings (e.g. new arrival items), they can still be recommended to users by

HTR if their topics are preferred by the users. As such, HTR effectively alleviated the

cold-start problem caused by dataset with high sparsity in user ratings (i.e. user ratings

cover only a small portion of all items).

Page 50

In the case of severe cold-start problem, HTR might have difficulties in making

quality recommendations, as it might be unable to allocate neighbourhoods for target

users based on their item preferences. CSHTR is specifically designed for such situations.

CSHTR finds target users’ neighbourhoods based on their taxonomic preferences instead

of item preferences, and hence it is capable of obtaining neighbours for target users who

have distinct tastes or few explicit ratings. Based on the neighbourhoods with similar

taxonomic preferences, CSHTR extracts the commonly preferred items from the

neighbours as candidate item lists. It then ranks and suggests these candidate items

according to the target users’ taxonomic preferences.

3.2.1 Notation

Before delving into algorithmic details, in this subsection we formally define the

concepts and entities involved in this research. These definitions will be also used in

subsequent chapters in this thesis and they can be tied easily to arbitrary application

domains.

Users , , … , . All users that have browsed items or

contributed item ratings in the sites are elements of . Possible identifiers

are globally unique names, user ids, URIs, etc.

Items (or Products) , , … , . All domain-relevant items are

stored in set . Possible unique item identifiers can be proprietary product

codes from an ecommerce site (e.g. Amazon.com’s ASINs) or globally

accepted code (e.g. ISBNs, ISSNs, etc.).

Implicit user ratings , , … , . Every user is assigned a set of

items that he or she has implicitly rated. Implicit ratings are

automatically inferred and collected from the user’s non-rating relevant

Page 51

actions (e.g. history of purchases, navigation history and product mentions),

therefore, they usually imply the users’ possible item interests rather than

clear indications of subjective item preferences (i.e. whether the users like or

dislike the items) (Montaner et al., 2003). Hence, usually indicates a set

of items being seen or interested by , and there are no precise values

associated with the items in to indicate the degree of like and dislike to

the items. Similarly, for these items \ that are not implicitly preferred by

, it can only be concluded that these items are unseen or uninterested by

(rather than disliked by ).

In general, implicit ratings are far more obtainable and accessible in

ecommerce sites and online communities than explicit ratings. Therefore,

when they are applied appropriately, implicit ratings can be a good means

for alleviating the cold-start problem (Schwab et al., 2000, Ziegler et al.,

2004).

Explicit user ratings , , … , . Every user is assigned a set of

items  that he or she has explicitly rated.

Explicit rating value , contributed by user to item . In

contrast to implicit ratings, explicit ratings are obtained by letting users to

judge items explicitly on a binary scale (e.g. classify an item as ‘like’ or

‘dislike’ or as ‘relevant’ or ‘irrelevant’) or discrete scale (e.g. rank an item

from 1 to 10, 1 indicates ‘dislike most’ and 10 indicates ‘like most’). In

order to express the degree of users’ item preferences in explicit ratings, we

use , to denote user ‘s explicit rating value to item .

Moreover, in order to accommodate different explicit rating scales, we

Page 52

assume the explicit ratings are normalised so that , 0,1 , where 0

indicates minimal satisfaction and 1 indicates maximum satisfaction.

User ratings , , … . All items that user has implicitly or

explicitly rated before, i.e. .

Taxonomic topic set , , … , . Set contains taxonomic topics

or categories for item classification. Each topic represents one

specific subject that items into which may fall. Topics express broad

or narrow concepts, when a topic’s concept is covered by (or is part of)

another, we call the former topic as sub-topic of the latter. We define map

: 2 retrieves all direct sub-topics for topics .

Based on the sub topic relation, we can define a strict partial order on the

topics in set to differentiate between super topics and sub-topics. Formally,

, , if , then is a subtopic of and there is a partial

order between and , denoted as . In addition, for simplicity,

we require that for all , , , so that one

topic can only have one direct super topic. With this requirement and the

map , we can recursively extract the taxonomy tree structure from the set

. Moreover, the same as all standard tree structures, the taxonomy tree has

exactly one top-most element, denoted as , with zero incoming-degree

representing the most general topic. In contrast, for the bottom-most

elements with zero outgoing-degree, they are denoted by and represent the

most specific topics.

An example of item taxonomy is shown in Figure 3.1. Within the item

taxonomy depicted in the figure, ‘ROOT (Books)’ is the root topic (i.e. )

covering the broadest concept, and ‘Apache’ and ‘Unix’ are the leaf topics

Page 53

(i.e. ) expressing the most specific concepts. The map returns the direct

sub-topics for any given topics in the taxonomy, for example,

Web Development Ecommerce,Web Design,Web Servers .

Item taxonomic descriptors , , … , . In order to describe

and classify items, every item is associated with a set of item

taxonomic descriptors . Note, an item can be described with multiple

descriptors because the item might possess a broad range of concepts,

strictly categorising the item under one single concept might be imprecise.

A taxonomic descriptor is a sequence of ordered taxonomic topics, denoted

by , , … , where ,  , and , , … , . The

topics within a descriptor are sequenced so that the former topics are super

topics of the latter topics, specifically, E and , where

0 . In our system, for any item descriptor  , , … , , it is

required that and .

Figure 3.2 shows an example list of items (i.e. books) with their

corresponding item taxonomic descriptors given under ‘Category’. For

example, the first book (‘Book#1’) in this list contains three item taxonomic

descriptors, and their corresponding leave topics (i.e. the most specific topics)

are ‘Apache’, ‘Network Administration’, ‘Network Programming’

respectively. With the defined information model, the item taxonomic

descriptors can be represented by "Book#1" , , , where:

"Books", "Computer & Internet", 

"Web Development", "Web Servers", "Apache"

"Books", "Computer & Internet", 

“Networking", "Network Administration"

Page 54

Books, Computer & Internet, 

“Networking", "Network Programming"

Figure 3.1: An example fragment of item taxonomy extracted from Amazon.com.

Page 55

Figure 3.2: An example list of items with their taxonomic descriptors.

3.2.2 Item Preferences based User Clusters

Clustering has been widely applied in recommender systems (especially

collaborative filtering based ones) to improve the computation efficiency (Cöster et al.,

2002, Sarwar et al., 2002, Jerome and Derek, 2004, Gui-Rong et al., 2005, Rashid et al.,

2006b, Rashid et al., 2006a). As mentioned in Section 2.1.2, collaborative filtering based

recommenders make recommendations to a target user by taking the opinions from other

users with similar item preferences to the target user. The process of finding users with

similar item preferences to the target user is commonly referred to as ‘Neighbourhood

Formation’. While neighbourhood formation is one of the most important steps in

making recommendations, it can also be the major performance bottleneck for

recommenders when the number of users and items in the system is large. The basic idea

behind clustering is to improve the online neighbourhood formation process by utilising

Page 56

offline computed user clusters. Figure 3.3 depicts how neighbourhood searching spaces

can be significantly reduced within the neighbourhood formation process based on the

pre-computed user clusters. Figure 3.3(a) shows that in standard collaborative filtering

recommenders, the target user’s profile (i.e. the circled dot) is compared with all other

user profiles in the dataset (i.e. all other dots within the dashed circle) in order to find the

top closest neighbours. In Figure 3.3(b), users are grouped into small clusters (i.e. dots

within the squares), hence, the searching space for forming the neighbourhood is reduced

within the target user’s cluster.

Figure 3.3: Reduce neighbourhood searching space with clustering

In order to form the neighbourhood for a given target user based on similarity of

users’ item preferences, a similarity measure is required to determine the degree of

similarity between two users’ item preferences. Pearson’s correlation coefficient and

cosine similarity count among the most prominent similarity measures for users’ item

preferences (Breese et al., 1998, Herlocker et al., 2002). In this thesis, Pearson

Page 57

correlation is adopted since it can accommodate the differences between users’ rating

styles (i.e. some users have a preference for the extreme values of the rating scale, while

others rarely deviate from the median) and, therefore, usually leads to better

recommendation quality (Herlocker et al., 2002, Herlocker et al., 2004, Jun et al., 2006).

The Pearson correlation coefficient measure used for computing the item preference

similarity between two users , is defined below:

,∑ , ,

∑ , ∑ ,

(3.1)

where is an item rated explicitly by both and , specifically

and , denote the average explicit ratings made by and . The average explicit

rating for a user can be computed by:

∑ ,

| |

Based on Equation (3.1), the user set can be divided into a set of user clusters

, , … , , such that and . For the sake

of convenience, let  denote the cluster that contains user . As the

clusters are constructed based on users’ item preference similarity, users within the same

cluster will have similar item preferences. There are many existing clustering techniques

which can be utilised for producing the user clusters, some widely recognised ones are k-

means, k-modes and x-means (Gui-Rong et al., 2005, Jain et al., 1999, Pelleg and Moore,

2000, Sarwar et al., 2002). Additionally, we have also developed an effective clustering

Page 58

method for recommender systems, HPC, and the detail of this technique can be found in

Appendix B.

3.2.3 Item Preferences - Taxonomic Preference Relation

Most recommender systems make recommendations by exploiting the relations

among users’ item preferences. For example, under the widely accepted assumption that

users must have similar tastes if they have similar item preferences (i.e. similar ratings to

the same items), given a set of items rated by a target user, collaborative filtering based

recommenders make recommendations by exploring other items that have been rated

similarly by the target user’s neighbours (Goldberg et al., 1992, Breese et al., 1998, Mira

and Dong-Sub, 2001, Lemire and Maclachlan, 2005). Recent studies on exploiting users’

taxonomic preferences to make recommendations are also based on an assumption that

users must have the similar content interests if they have similar taxonomic preferences

(Sollenborn and Funk, 2002, Ziegler et al., 2004, Middleton et al., 2002). Intuitively and

based on our observations, in this thesis, we propose the following assumption about the

relation between users’ item preferences:

Assumption 3.1. (Item Preferences - Taxonomic Preference Relation) Users who are

in the same item preference based neighbourhood or cluster share not only similar item

preferences but also similar taxonomic preferences.

In the case of clustering based neighbourhood formation, Assumption 3.1

suggests that the users within one cluster should have apparent similar taxonomic

focuses and the taxonomic focuses of the users in different clusters should be different.

The proposed HTR and CSHTR techniques in this thesis are designed and implemented

based on this assumption. Through our experiments in Section 3.3, we have shown that

Page 59

the proposed techniques have gained significant improvements in recommendation

making in both normal and cold-start environments. Moreover, the validity of

Assumption 3.1 has been verified empirically with the use of information gain measure.

The detailed verification and experiment process is detailed in Section 3.3.2.

3.2.4 Extraction of User’s Taxonomic Preferences

In this section, the techniques we employed to extract users’ taxonomic

preferences are described. In our thesis, users’ taxonomic preferences are considered in

two different aspects, namely ‘personal taxonomic preference’ and ‘cluster taxonomic

preferences’, which are discussed in detail in the following subsections.

3.2.4.1 Personal Taxonomic Preference

A user’s personal taxonomic preference implies the taxonomic topics that the

user has shown his or her interests to in the past. We capture a user’s personal taxonomic

preferences through examining the taxonomic topics from the items that the user has

rated (both implicitly and explicitly) before. In this thesis, because the taxonomic topics

are contained in a taxonomic tree structure and impose a hierarchical relation among

each other (see Section 3.2.1), therefore, the following factors can be considered when

designing a technique to compute users’ personal taxonomic preferences:

Page 60

1) Frequency of a user’s topic interest indication. When two topics

, are in the same level (e.g. if they are both leaf topics), a user

may be more interested in than , if the user has rated more items

belonging to than items belonging to . For example, suppose a user

has browsed the three books in Figure 3.2, he or she might be more

interested in the topic ‘Apache’ than in ‘Java’, because all of these three

books are related to ‘Apache’ and only ‘Book#3’ is related to ‘Java’.

2) Item taxonomic topic hierarchy. When two topics , have the

same frequencies in a user’s item ratings, the user may be more interested

in than , if is a sub-topic of . For example, if a user has only

rated ‘Book#2’ in Figure 3.2, he or she might be more interested in topic

‘Apache’ than ‘Web Servers’ even the frequency of the two topics rated

by the user are the same. It is because there are many other topics under

‘Web Servers’ such as ‘Unix’, ‘Linux’, ‘Windows’, etc., and the topic

‘Apache’ describes a more specific domain concept that is encompassed

by ‘Web Servers’. The user who rated ‘Book#2’ might only be interested

in ‘Apache’ rather than ‘Unix’ and ‘Linux’, hence items belonging to

‘Apache’ would have more chances to be preferred over items only

belonging to ‘Web Servers’.

3) Topic concept coverage. Given two sibling topics , (i.e. is

not ’s super topic, and vice versa), has a broader concept coverage

than if contains more sub-topics than . If and have the

same occurrence frequency in a user’s item ratings, the user might be

more interested in than , as has a narrow coverage and more

likely contains the topics that the user prefers.

Page 61

4) Relevance of concepts in sibling topics. Sibling topics must have

something in common that is captured by their super topics. Given two

sibling topics , , their common features can be observed through

their super topics. If a user has interests in , it is reasonable to think

that the user might also be interested in since and share some

common features.

Ziegler et al. (2004) have proposed a technique to derive user personal

taxonomic preferences from user implicit ratings. We have thoroughly analysed

Ziegler’s technique and found that it has taken all the four factors mentioned above into

consideration. In our work, we adopted Ziegler’s technique to generate user personal

taxonomic preferences. For an user , the user’s personal taxonomic preference

can be modelled by a | |-entry vector, called personal taxonomic profile vector, denoted

as , , , , … , ,| | . Each entry , in represents the degree of ’s

preference or interest (i.e. personal taxonomic preference score) to the topic in (i.e.

). In order to measure the similarity between two profile vectors, user-wise

normalisation is applied, such that:

:  ,

| |

where is the normalisation factor that can be any positive number, in this thesis we set

| |.

If item taxonomic data is available, users’ personal taxonomic profile vectors can

be generated from users’ ratings to items because each item is associated with a set of

taxonomic topics. In this thesis, the taxonomic topics related to an item can be obtained

from the item’s descriptors. For the items rated by a user, their descriptors need to

equally contribute scores to the user’s personal taxonomic profile vector. Specifically,

Page 62

for any item rated by user (i.e. ), the score contribution of any ’s

descriptor to ’s profile vector can be computed by:

,| | | |

where | | is the number of items rated by and | | is the number of topic

descriptors of item .

As , is meant to be distributed to topics , , … , in descriptor ,

therefore, it is required that:

  ,   

(3.2)

where is the score assigned to topic , and it can be computed by:

 1  

(3.3)

where returns the number of topic ’s siblings. In other words,

1 resolves the number of ’s immediate children or sub-topics. is a

propagation factor that permits fine-tuning for the significance of topic specificity and

depth in the profile construction process.

It can be observed in Equation (3.3) , is inversely proportional to the

number of ’s direct sub-topics, and, therefore, may be assigned with a higher

score if it covers a more specific domain concept (i.e. in accordance with the third factor

Page 63

described at the beginning of this section). It can also be observed from the equation that

for any 0 , (when 0 1 ), this allows a

hierarchical score decay from the most specific topic (i.e. ) to the most general topic

(i.e. ). Thus, the second factor (i.e. item taxonomic topic hierarchy) is also satisfied.

Moreover, because the final score , for a topic in a profile vector is

accumulated with multiple that are computed from the descriptors of items rated

by , the first design factor is satisfied. Finally, it can be easily observed that the fourth

factor is also satisfied with this approach, because two descriptors with different leaf

topics might still share common intermediate topics and these common topics might be

assigned with similar topic scores.

A brief example of the taxonomic score computation described above is shown

below:

Example 3.1 Suppose a user has rated the three books listed in Figure 3.2, and

these books are categorised based on the book taxonomy depicted in Figure 3.1. Based

on Equation (3.2) and (3.3), we can calculate how scores are distributed to the topic

entries of ’s personal taxonomic profile vector from any given item’s descriptor. In

this example, we demonstrated how scores are distributed through the first descriptor of

‘Book#1’, that is,

"Books", "Computer & Internet", 

"Web Development", "Web Servers", "Apache"

Suppose that 900 defines the normalisation factor for the profile vectors, then the

score assigned to any one of descriptors of ‘Book#1’ amounts to:

,| | | “Book#1” |

9003 3

100

Page 64

Next, as the exact score value for the leaf topic ‘Apache’ is unknown, we let

"Apache" . Based on Equation (3.3), the relative score value for the parent

topic of ‘Apache’ (i.e. ‘Web Servers’) can be computed (assuming the propagation

factor 1):

"Web Servers"  "Apache"

"Apache" 11

1 1 2 

Similarly, the score for the topic ‘Web Development’ can be computed based upon its

parent topic (i.e. "Web Servers" ):

"Web Development"  "Web Servers"

"Web Servers" 11 2

2 1 6 

Accordingly,

"Computer & Internet"24 

"Books"  96 

Next, based on Equation (3.2), the exact topic scores can be computed by solving:

"Apache" "Web Servers" "Web Development"

"Computer & Internet" "Books" ,  

Thus, 

2 6 24 96100 

58.18 

Finally, by applying the exact value of   to the topics, we can obtain: 

"Apache" 58.18

"Web Servers"2

29.09

"Web Development"6

9.70

Page 65

"Computer & Internet"24

2.42

"Books"96

0.61

In Ziegler’s research, the taxonomic profile vectors are compared with each

other to measure the taxonomic preference similarity between two users (therefore, the

user-wise normalisation is required). In contrast, this thesis emphasises the extraction of

these user preferred taxonomic topics using Zeigler’s method, and then merge the

personal topic scores with the cluster taxonomic topic scores (to be discussed in Section

3.2.4.2) to generate a more comprehensive user taxonomic preference profile. In order to

uniformly merge personal topic scores with cluster topic scores, we further normalise the

topic scores obtained from the taxonomic profile vectors with min-max normalisation

technique, and use the normalised topic scores as the personal taxonomic topic scores in

the final personal taxonomic profile vector. Specifically, let , be the topic score to the

topic in a user ’s personal taxonomic profile vector computed using

Ziegler’s method, the user’s final topic preference score to the topic in C can be

obtained by:

_ ,   , minmax min

(3.4)

where min and max are the minimal and maximal score values in ’s

taxonomic profile vector respectively. After the normalisation, user ’s most

preferred topic will receive the topic score _ , 1 , and the most

disliked topic will receive _ , 0.

Page 66

3.2.4.2 Cluster Taxonomic Preference

As described in Section 3.2.4.1, the personal taxonomic topic preference score

_ , is obtained from converting users’ rating data and can be used to measure

users’ personal topic interests. In order to obtain a more comprehensive profile about a

user’s taxonomic interests, it is necessary to study the taxonomic interests of other users

within the same cluster, group or neighbourhood, since users within the same cluster

usually share similar taxonomic interests as assumed in Assumption 3.1. Therefore, we

propose to estimate the user’s taxonomic preference by combining his or her personal

taxonomic preferences with the cluster taxonomic preferences.

In order to extract the cluster level taxonomic preferences, we firstly build a

cluster-based taxonomy similar to the global taxonomy defined in Section 3.2.1 (i.e.  )

for each cluster . Formally, we define the cluster-based topic set:

| , , ,  

Further, a corresponding map for topics , such that

extracts the direct sub-topics of . Note, because items that have been rated in one

cluster might not be rated at all in other clusters, therefore, it is possible that the cluster-

based topic set contains only a subset of all topics, specifically . Furthermore,

from , we can also conclude that   .

Based on the local cluster taxonomy tree, we can measure the distinctness of a

topic within a local cluster in accordance with the global user set. With the

distinctness, we can determine how popular a topic is in a cluster and how unique a topic

is to a cluster comparing to other clusters. The distinctness can be assessed by the

following equation:

Page 67

_ ,0, _ ,

_ ,_ ,

,   

(3.5)

where _ , is the number of user ratings involving taxonomy topic within a

given user set . Specifically,

_ , _ ,  

where _ , checks if item belongs to topic ,

_ ,1, | ,

0,

Moreover, in Equation (3.5) is a user defined constant, it is used to filter out

topics that are not popularly interested by users. For example, when is set to 50, a

topic needs to be involved in more than 50 item ratings in order to be considered

important within a given user cluster.

It can be easily observed from Equation (3.5), the higher the computed topic

score _ , , the higher the possibility the taxonomy topic is unique and

popular in cluster . It is because in order to obtain a high value for _ , ,

the value of _ , need to be larger than the given threshold and approach

to _ , . It implies that the topic not only need to be popular for the whole

user set (i.e. ) but also within only one cluster . In contrast, if a topic has high

popularities in multiple clusters (i.e. high _ , values for different clusters),

then it will receive a low value for _ , as _ , will be much

larger than _ , ; it indicates that the topic is not unique in cluster .

Page 68

In order to linearly combine the personal and cluster level taxonomic preference

scores (i.e. _ , and _ , ) together, we also normalise

_ , with min-max normalisation. The normalised cluster level taxonomy

preference score is denoted by _ , , specifically,

_ ,_ , min_sc

max_sc  min_sc 

(3.6)

where

min_sc min _ ,

max_sc max _ ,

3.2.4.3 Merge Personal and Cluster Taxonomic Preferences

In this thesis, a user’s taxonomic preference is constructed with respect to both

the personal taxonomic interests (as described in Section 3.2.4.1) and the cluster level

taxonomic interests (as described in Section 3.2.4.2). In the aspect of personal taxonomic

interests, the user’s topic interests are investigated. In the aspect of cluster taxonomic

interests, we induce the topics that might be potentially interested by the user through

exploring the taxonomic topic interests of the user’s neighbourhood. Having obtained the

user’s personal taxonomic preference profile and the taxonomic preference profile of the

user’s cluster, we compute the user’s taxonomic preferences by linearly combining the

two profiles. Formally, for user and topic the user’s taxonomic preference

score to is:

Page 69

_ ,   _ , 1 _ ,  

(3.7)

where 0 1 is a user controlled parameter for adjusting the weight between the

personal level and cluster level taxonomic preferences in the final taxonomic preference

score computation.

3.2.5 Hybrid Taxonomy Recommender

In this section, we describe the proposed Hybrid Taxonomy Recommender

(HTR) that incorporates users’ taxonomic preference profiles described in Section 3.2.4

with the item-based collaborative filtering (item-based CF) to improve recommendation

quality.

HTR generates item recommendations by combining the estimates to item

preferences and the estimates to taxonomy preferences. In this section, we firstly explain

the item-based CF technique that has been applied in HTR for item preferences

estimation, then the method to calculate users’ taxonomic preferences, finally the

algorithm to generate a list of recommended items. Item-based CF recommends item

to user based on the item similarity between and the items that have been

rated by . The similarity between two items , is computed based on checking

whether these two items are rated similarly by users (Badrul et al., 2001). Specifically:

_ ,∑ , ,

∑ , ∑ ,

(3.8)

Page 70

where , and , denote user ’s rating to item and respectively. and are

the average ratings of and . is the set of users who have rated both and .

is defined as:

| ,

Note, it is possible that two items are never rated together by a user, i.e. .

In such case, _ , returns a special value that is a label indicating ‘Not

Computable’.

As mentioned above, the preference estimate to item for user is based

on the similarities between and the items \ rated by . In order to achieve

it, we need to find items that are explicitly rated by target user and are computable

with the target item   . That is,

, \ | _ ,

Finally, the item preference prediction to item for user can be

computed by:

,

∑ _ , ,,

∑ | _ , |, 

(3.9)

As rating values (e.g. , ) are pre-normalised between 0 and 1 as described in

Section 3.2.1, it can be easily observed from Equation (3.9) that the value range of item

preference prediction score (i.e. , ) is also between 0 and 1. When , is close to 1, it

indicates might highly prefer . In contrast, when , is close to 0, it indicates that

might have no interests in .

Page 71

Users’ taxonomic preferences are predicted based on the topic scores computed

by Equation (3.7). Let , denote the prediction of a user ’s taxonomic preference to

item  , it can be computed as below:

, _ ,  

(3.10)

where | , is the set of item ’s topics.

The recommendation of an item to a user will be determined according to both

the user’s item preference prediction computed using Equation (3.9) and the user’s

taxonomic preference prediction computed using Equation (3.10). In order to

recommend a set of items to a target user , we firstly form a candidate item list

containing all items rated by ’s neighbours (i.e. ) but not yet rated by .

Next, for each item in the candidate list, we compute the item preference score and

taxonomic preference score for the item. The proposed preference score for each

candidate item can then be computed by combining the item preference score ( , ) and

the item taxonomic preference score ( , ) together. Finally, candidate items with

highest preference scores are recommended to the user  , and these recommended

items are sorted by their corresponding score values. The complete algorithm is listed

below:

Page 72

Algorithm 3.1 _ ,

Input is a given target user

is the number of items to be recommended

Output a list of items recommended for

1) SET \ , the candidate item list

2) FOR EACH

3) SET , , 1 ,

4) END FOR

5) Return the top k items with highest , scores to .

From line (3) of Algorithm 3.1, it can be observed that the predicted ranking score for an

item is computed by a linear combination of item preference score , and topic

preference score , . The coefficient , computed by Equation (3.11) below is used to

adjust the weight between of , and , :

1 1 

(3.11)

where is the ratio between the number of the items that are commonly rated with item

by and other users and the number of the items rated by . Specifically,

| , |

| |

and is a user controlled variable that allows manual adjustment for the weight between

of , and , , such that 0,1 .

Page 73

Figure 3.4. The impact of different values on ( 0.28)

Figure 3.5. The impact of different values on

0

0.2

0.4

0.6

0.8

1

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

α2

ω

0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

α2

ω

θ 0.01θ 0.25θ 0.5θ 0.75θ 0.99

Page 74

In Equation (3.11), reflects the confidence to the quality of , , because the

more the target user’s past rated items related to the target item, the higher the accuracy

of the item preference prediction (i.e. , ) will be. When increases will increase

too, thus, , will receive higher weight in the final score (i.e. , ). The relationship

between and with five different settings is depicted in Figure 3.5. It can be

observed from the figure that and are proportional to each other.

Variable , on the other hand, allows manual adjustment of , thus, if is large

(e.g. 0.9) , will still receive high weight even is small. Figure 3.4 and Figure 3.5

demonstrates how can be used to control the value of . When we consider item

preference and item taxonomic preference are equally important in recommendation

making, we can set 0.5. In such a case, as can be easily observed in Equation (3.11),

Figure 3.4 and Figure 3.5, the value of will entirely depend on (i.e. ). The

use of the control variable in the design of the algorithm enables incorporation of

subjective considerations in adjusting the weights of , and , . For example, for

application domains where users’ considerations in item quality outweigh item topic

relevance, can be set to a higher value to allow recommendations mainly depend on

users’ item preferences.

The value (and thus ) is also a good indicator for identifying cold-start

situations. A high value indicates that there exist many users who have commonly

rated the item with the target user, hence there is no cold-start problems. A low or zero

value indicates that there is very few or none users who have commonly rated the item

with the target user, which is one of the cold-start problems.

The value of is automatically adjusted along with the change of the number of

users who commonly rated a given item . The higher the value of , the more the

users who commonly rated the item are (i.e. a normal situation without severe cold-start

Page 75

problems) and, thus, the item preference , estimated based on these users’ rating data

becomes more important and reliable. In this case, the predicted item preference ,

makes a larger contributions to the overall prediction score , than the contribution

made by the predicted taxonomic preference , . On the other hand, if the value of

is low (i.e. a cold-start situation), the taxonomic preference prediction becomes more

important and will contribute more to the overall prediction score , than does the

predicted item preference. This design ensures that taxonomic preferences are used to

supplement or enrich the item preference prediction, especially in cold-start situations.

3.2.6 Cold-Start Proof Hybrid Taxonomy Recommender

By utilising users’ taxonomic preferences, the HTR technique proposed in

Section 3.2.5 is effective even when there is only a small amount of users sharing similar

explicit item ratings to the target user (i.e. when ω is small). However, the proposed

algorithm requires that the given target user can be correctly allocated into one of the

pre-computed user clusters (i.e. ) based on the explicit ratings. Hence,

in severe cold-start situations where the given target user has very distinct tastes and

cannot be allocated to any of the clusters (i.e. ) or his or her explicitly

rated items have not been rated by more than one previous user, the proposed HTR

technique in Section 3.2.5 suffers from the severe information shortage and cannot

make satisfactory recommendations. In this section, we propose another technique,

namely Cold-Start Proof Hybrid Taxonomy Recommender (CSHTR), specifically

designed for making recommendations in the severe cold-start situations. In Section

3.2.3, we suggest that a group of users with similar item preferences might share similar

taxonomic preferences. While the proposed HTR technique applies this rule to discover

Page 76

target users’ potential taxonomy interests for improving recommendation qualities,

CSHTR utilises this rule from the other direction by using the target users’ taxonomic

preferences to discover their potential item preferences. More specifically, instead of

allocating a given target user to a user cluster based on explicit item

preferences, CSHTR finds ’s belonging cluster by comparing ’s taxonomic

preferences with each user cluster’s general taxonomic preferences. For this purpose, we

need to generate users’ taxonomic preferences and each cluster’s general taxonomic

preferences.

The taxonomy vector , , , , … , ,| | as described in Section 3.2.4.1

will be used to represent a user ’s personal taxonomy preferences. The general

taxonomic preferences of a cluster can be obtained by computing the mean

vector of all users’ taxonomy vectors within . Specifically, the taxonomy vector for a

cluster is denoted as:

, , , , … , ,| |

where

,∑ ,

| |

Note, as these taxonomy vectors are mainly used for similarity comparison,

therefore it is not necessary to further normalise these vectors as described in Section

3.2.4.1.

With taxonomy vectors, the taxonomic preference similarities between two

users or between a user and a user cluster can be computed with cosine similarity

measure:

Page 77

_ ,| || |

(3.12)

Based on (3.12) we can find the user cluster that has the most similar taxonomic

preferences with a given user by:

_ argmax   _ ,  

(3.13)

In severe cold-start situations, based on the taxonomy preference similarity

discussed above, the target user can still be located to a user cluster even

. There are three reasons behind it:

In most cases, the number of taxonomic topics is much smaller than the

number of items (i.e. | | | |). Therefore, the possibility of common

entries in taxonomic topic vectors is much higher than that in item vectors.

Multiple different items might share common topics. For a user who only

rated new items that no one has rated before, it is still possible to find users

with similar taxonomic interests to the user, because there might still be

many items with similar topics to these new items.

Taxonomic topics are organised in a hierarchical tree structure, and impose

hierarchical relations on each other; hence, different topics may be covered

by common super topics. For a user who is interested in a new topic that no

one has known yet, it is still possible to locate the user’s neighbours by

finding users with interests to the super topics of this new topic.

Page 78

However, Equation (3.9) becomes unsuitable for generating item preference for

the target user because the target user may not have any items commonly rated by

previous users, i.e. , . For the severe cold-start situations, we propose

to compute the commonly preferred items within the user cluster and treat these

commonly preferred items as the item preferences of each user in this cluster. A

commonly preferred item can be determined by the popularity of the item in the cluster

and the average of the item’s explicit rating scores given by the users in the cluster who

rated the item. Specifically, the degree of general preference to an item by the

users in a cluster can be computed by:

, , 1 ,  

(3.14)

where , is the average explicit ratings to and , measures ’s

popularity in which are computed by:

,∑ ,

| | 

(3.15)

,1, | |

| |,   

3.16  

In Equation (3.15) and (3.16), denotes the set of users in who rated

explicitly, that is:

Page 79

The popularity of an item in a cluster is measured by the number of users in this

cluster who rated the item, the more users who rated the item, the higher the popularity

of the item is. For easy description, we call the number of users who rated an item in a

cluster the Population Value of the item in this cluster. In Equation (3.16), we designed

an upper bound for normalising the popularity score so that , 0,1 . The

upper bound is computed by utilising the common 95% empirical rule (Tabachnick

and Fidell, 2006):

2

where is the average population value of the items in , which is,

∑ | || |

and is the standard deviation of the population values of all items in ,

1| |

In both and , denotes the set of all items being rated by users

in cluster , specifically:

In this thesis, it is assumed that is normally distributed (i.e. under

normal distribution). Therefore, the empirical rule based upper bound allows

approximately 95% items in ’s candidate item list (i.e. ) have a smaller population

value (i.e. ) than . If | | , it means that item is popularly rated and,

therefore, preferred by the users in . In this case we set 1 to , . This design

ensures the value of is set with a reasonable value, so that , can be reasonably

distributed between 0 and 1. That is, when is set too large, most of , values

Page 80

will be very small; when is set too small, most of , values will equals to 1.

Furthermore, one might suggest that setting the upper bound to the maximum value of

| | might be sensible solution, that is:

max | |

However, it is very likely the maximum | | may be in fact an outlier, and

consequently resulting very small , values.

In Equation (3.14), 0,1 is a user controlled variable for adjusting the

weights between the average item preference and item popularity.

Overall, in the severe cold-start situation, a target user’s preference to a given

item is predicted based on the general preference to the item in the user’s belonging

cluster and the taxonomy similarity between the target user and the item. The detailed

CSHTR algorithm is listed below:

Algorithm 3.2 _ _ ,

Input is a given target user

is the number of items to be recommended

Output a list of items recommended for

1) SET _ \ , the candidate item list

2) FOR EACH

3) SET _

4) SET , , 1 _ ,

5) END FOR

6) Return the top items with highest , scores to .

Page 81

In the line (4) of the algorithm, _ , computes the similarity between

user ’s taxonomy vector (i.e. ) and item ’s taxonomy vector (i.e. ). Item ’s

taxonomy vector can be formed by assuming it as the taxonomic profile vector for a

user who only rated item , that is:

where is the taxonomic profile vector for a dummy user , such that .

0 1 is a user controlled variable for adjusting the weights between the

predicted item preference (i.e. , ) and predicted taxonomic preferences (i.e.

_ , ).

Insufficiency of rating data is one important reason resulting in the cold-start

problem. As the proposed CSHTR technique determines user neighbourhoods only

based on taxonomic data, and makes recommendations to the target user based on the

commonly preferred items no matter whether the target user has rated or not rated these

items, the insufficiency of explicit rating data is not crucial for CSHTR to make

recommendations. Thus, CSHTR is capable of generating quality recommendations in

severe cold-start situations. Moreover, unlike the TPR technique proposed by Ziegler et

al. (2004) which makes recommendations only based on taxonomic preferences, the

proposed CSHTR incorporates both item preferences computed from commonly

preferred items and taxonomic preferences together, therefore, yields better

recommendation quality.

3.3 EXPERIMENTS AND EVALUATION

The following sections present experimental results that were obtained from

evaluating our approach. In Section 3.3.1, the dataset we employed for the experiments

Page 82

is discussed. In Section 3.3.2, the suggested assumption about the relationship between

item preference and taxonomic preference (see Section 3.2.3) is verified based on the

information gain technique. Finally, in Section 3.3.3 the proposed HTR and CSHTR

techniques are empirically evaluated.

3.3.1 Data Acquisition

In this thesis, the ‘Book-Crossing’ dataset (http://www.informatik.uni-

freiburg.de/~cziegler/BX/) is chosen to conduct the experiments. The ‘Book-Crossing’

dataset is collected by Cai-Nicolas Ziegler in a 4-week crawl (August / September 2004)

from the Book-Crossing community (http://www.bookcrossing.com/) with kind

permission from Ron Hornbaker, CTO of Humankind Systems. It contains 278,858

users (anonymised but with demographic information) providing 1,149,780 ratings

(explicit / implicit) about 271,379 books. In the user ratings, 433,671 of them are the

explicit user ratings, and the rest of 716,109 ratings are implicit ratings.

The book taxonomy and book descriptors for the experiments are obtained from

Amazon.com. Amazon.com’s book classification taxonomy is tree-structured (i.e.

limited to ‘single inheritance’) and, therefore, is perfectly suitable to the proposed

technique. The average number of descriptors per book is around 3.15, and the

taxonomy tree formed by these descriptors contains 10,746 unique topics.

3.3.2 Verification for Item Preferences - Taxonomic Preference Relation

The assumption we proposed in Section 3.2.3 suggests that the users within one

cluster should have apparent similar taxonomic focus and the taxonomic focuses of the

users in different clusters should be different. In this section, we use information entropy

Page 83

to measure the certainty of user clusters’ taxonomy focuses and empirically validate the

proposed assumption by using information gain measure.

Information gain is commonly used in decision tree construction (Russell and

Norvig, 2002) to measure the increase or decrease in the outcome certainty when

dividing data with a given attribute. When the information gain is high, it indicates that

the divided datasets are more certain about some features. In the case of user clusters

discussed in this chapter, high information gain indicates that the certainty of the

taxonomic focuses of user clusters is high. By adopting the information gain measure,

we can investigate whether different clusters have apparent taxonomic focuses and the

taxonomic focuses are different in different user clusters. The information gain can be

calculated as below:

Pr

(3.17)

where Pr  is the probability that an item rating is made by an user in cluster , that

is,

Pr∑ | |

∑ | |

denotes the information entropy for a given user space. The concept of

information entropy is adopted in this thesis to measure the degree of taxonomic focus in

a user set (i.e. a cluster or a neighbourhood). If the information entropy is high for a user

set, then there is no apparent taxonomic focuses in the set (i.e. users in the set prefer all

taxonomy topics equally). In contrast, if the information entropy is low, then it indicates

Page 84

certain topics are popularly preferred within the user set. The information entropy

formula is depicted below:

Pr  , Pr  ,  

(3.18)

where    denotes all leaf topics in  , that is: 

|  

and Pr  ,  denotes  the  probability  that  the  users  in  the  user  set      are 

interested in the taxonomy topic  , specifically: 

Pr ,∑ | , |

∑ ∑ | , |

where , is the set of items that are rated by and can be categorised by topic ,

specifically:

, | , ,

  For  a  given  clustering  , , … , ,  if  all   are  low 

,  then  it  means  the  taxonomic  focuses  are  apparent  in  all  clusters  

, according to Equation (3.17), the information gain is high. 

Based on the experiment dataset described in Section 3.3.1, we extracted 10,000

users with more than 10 explicit past ratings (i.e. 10) from the 278,858 users in

the entire dataset. k-means clustering technique is then applied to divide these 10,000

users in the dataset into 100 user clusters according to their explicit ratings (detailed

information for user clustering is described in Section 3.2.2). We have tried to produce

different number of clusters for the dataset (i.e. different values for k), and we have

found by setting k to 100 (i.e. 100 clusters) can produce clusters with reasonable qualities.

Page 85

In order to form the baseline of our experiment, we also constructed 100

randomly formed user clusters from the same user set. The population distribution of the

randomly formed user clustering partition is similar to the target clustering partition.

That is,

, | | | | 

where is the target clustering partition generated by k-means, and is

the randomly formed partition. For , users within the same cluster have similar

item preferences, in contrast, users within the same cluster of have no apparent

item preference similarities among each other.

Our first experiment is to show if user clusters have stronger taxonomic focuses

than the entire dataset when only explicit ratings are considered. It is shown in the first

column of Table 3.1 the result information gain is 0.823, which is a big increase when

comparing it with the information gain obtained from the randomly formed cluster

partitions (i.e. -0.385). This result shows that, by clustering users with their explicit

ratings, each user cluster has its own taxonomic focuses.

Since our clusters are generated based on only explicit ratings, it might be unfair

if we only consider explicit ratings in calculating taxonomy information gain. Hence, we

further include the implicit ratings in computing taxonomy information gain. With

identical cluster settings, we still get a strong information gain increase (i.e. 0.458) when

comparing to the information gain obtained from the random formed clusters (i.e. -

0.319). Based on the information gain analysis, we can conclude that users within 

the  same  clusters  not  only  share  similar  item  preferences,  but  they  also  share 

similar taxonomic preferences.  

Page 86

Table 3.1. The effect of user clustering on taxonomy information gain

Explicit Ratings Explicit + Implicit

Ratings

users clusters formed based on user ratings ( )

0.823 0.458

Randomly formed user clusters ( )

-0.385 -0.319

3.3.3 System Evaluations

In this section, the computation efficiency and recommendation quality of the

proposed HTR and CSHTR techniques are empirically evaluated.

3.3.3.1 Experiment Framework

In this section, the underling system framework employed for conducting the

experiments is described.

All recommenders being used in the experiment are developed using the Taste

(http://taste.sourceforge.net/) framework that is popularly used for evaluation in

recommender research community. Taste provides a set of standardised components for

developing recommenders, and therefore it ensures the comparability of the developed

recommenders fairly. Moreover, Taste also provides an evaluation framework allowing

researchers or developers to evaluate the performances of their recommenders with a

standardised test bed easily and effectively.

Including the proposed HTR and CSHTR techniques, we have constructed eight

different recommenders in total for the experiments. These recommenders are:

Page 87

Item-based Recommender (IR) Standard item-based collaborative filtering

recommender, the detailed algorithm is given in (Badrul et al., 2001,

Deshpande and Karypis, 2004). This recommender is constructed by

employing the default implementation from the Taste framework, therefore,

the validity of the experiment results is further ensured.

In general, IR computes item preference scores (i.e. , ) for a target user

to all items based on Equation (3.9), and recommends top k

items with the highest item preference scores to . Note, IR only uses

explicit ratings for its recommendation making, and hence implicit rating

data are discarded.

Item-based Recommender with User Clustering (IRC) The item preference

prediction of this recommender is the same as IR, and the only difference

between IRC and IR is that IRC optimises its computation efficiency by

utilising the pre-computed user clusters. More specifically, while IR needs to

compute item preference scores for all items when making a

recommendation for a user , IRC only needs to compute the scores

for items that have been rated in ’s user cluster (i.e.

) .

Slop One Recommender (SO) A well known modern item-based

recommendation technique (Lemire and Maclachlan, 2005), it features on its

implementation simplicity and computation efficiency. The implementation

of this recommender is provided by the Taste framework, so the validity and

accuracy of the implementation is ensured. This recommender utilises only

explicit ratings in its recommendation making process as similar to IR and

IRC. The reason for including SO in the experiments is to ensure that the

Page 88

general recommendation performance achieved with only explicit rating

data can be objectively observed.

The three recommenders mentioned above are existing standard

recommender models. They serve as the benchmark models for this

evaluation.

Taxonomy Product Recommender (TPR) A taxonomy-based

recommender proposed by Ziegler et al. (2004). This study uses similar

taxonomy scheme to our thesis, and therefore can be a good benchmark. For

more details about the TPR, please refer to Section 2.2 and (Ziegler et al.,

2004). Note, TPR uses only implicit rating data for its recommendation

making, and hence explicit rating data are discarded.

Item-based Recommender with TPR (ITR) The combination of the item-

based CF (i.e. IR) and TPR. The hybridisation scheme is identical to HTR

(see Algorithm 3.1). The only difference is that , is computed using

Ziegler’s method (i.e. TPR). As ITR is a hybrid of IR and TPR, therefore, it

utilises both explicit and implicit rating data for its recommendation making

process.

ITR is included in the experiment to allow the proposed HTR technique to

be objectively and fairly evaluated by comparing with the ITR. It is because

both HTR and ITR use explicit and also implicit rating data, while IR, IRC

and SO uses only explicit rating data and TPR uses only implicit ratings

which might make (the comparison between HTR with IR, IRC, SO and

TPR lack of fairness.

Page 89

Hybrid Taxonomy Recommender (HTR) The proposed HTR method as

described in Section 3.2.5, it uses both users’ explicit rating data and implicit

rating data for recommendation making.

Cold-Start Proof Hybrid Taxonomy Recommender (CSHTR) The

proposed CSHTR method as described in Section 3.2.6, it is mainly

designed for severe cold-start situations. CSHTR uses both users’ explicit

rating data and implicit rating data for recommendation making.

Hybrid Taxonomy Recommender (with only explicit ratings) (HTR_E)

The proposed HTR method using only explicit ratings. The purpose of

including this recommender in the experiments is to ensure fair comparison

with IR, IRC, SO, which use only explicit ratings.

3.3.3.2 Parameterisation

In this section, the parameter values we assigned to configure the HTR and

CSHTR techniques for the experiments are detailed.

For the configuration of HTR:

The propagation factor for Equation (3.3) is set to 0.75. This setting

confirms to the configuration suggested by Ziegler et al (2004), and

therefore it ensures the experiment results from our study and (Ziegler et al.,

2004) can be compared. Assigning with values less than 1.0 allows higher

scores to be assigned to the super topics in the taxonomic profile vectors,

and it in turn allows profile vectors with similar scores in their super topic

entries to be considered closer. Amazon.com’s item taxonomy is deeply

nested and topics tend to have many siblings, therefore, topics in higher

levels (i.e. super topics) tend to have very small score values. By setting

Page 90

with a smaller value (i.e. 0.75), we ensure the score distributions for the

profile vectors constructed from Amazon.com’s item taxonomy are more

sensible.

The filter parameter for Equation (3.5) is set to 50. Therefore, a topic

needs to be involved in more than 50 item ratings in order to be considered

important within a given user cluster (i.e. receive _ , score

with value larger than 0).

The adjustment parameter for Equation (3.7) is set to 0.4. This setting put

slightly higher emphasis on the personal level taxonomic preferences for the

final item taxonomic preference score computation (i.e. _ , )

(therefore, less emphasis on the cluster level taxonomic preferences).

The adjustment parameter for Equation (3.11) is set to 0.6. This setting put

slightly higher emphasis on item preferences than item taxonomic

preference for the final recommendation ranking score computation (i.e.

, in Algorithm 3.1).

For the configuration of CSHTR:

The adjustment parameter for Equation (3.14) is set to 0.7. Therefore, in

the computation of cluster ‘s general preference to item (i.e.

cpref uc, t ), there are more emphasis on ’s average item preference to

item than ’s popularity in .

The adjustment parameter for the final recommendation ranking score

computation (i.e. , in Algorithm 3.2) is set to 0.5. This setting put the

same emphasises on both predicted item preference (i.e. , ) and

taxonomic preferences (i.e. _ , ).

Page 91

3.3.3.3 Evaluation Metrics

For the recommendation quality evaluation, we randomly divided each user

‘s past ratings (i.e. ) into two parts, one for training and another for testing. We

use to denote ‘s training rating data and to denote the testing rating data, such

that , , and | | | |. The testing data actually consists of

three types of items, and they are:

Items implicitly rated by :

Items preferred by : | , , , is the average

rating of user ’s explicit ratings.

Items not preferred by : \ 

In the experiment, the recommenders recommend a list of items to based

on the training set , and the recommendation list will then be evaluated with or .

There are two objectives in this experiment. The first objective is to evaluate

whether a recommender’s performance can be improved by incorporating the item

taxonomy information and the suggested assumption for the relation between item

preference and item taxonomic preference into the recommendation making process (i.e.

whether the proposed HTR technique outperforms other techniques). The second

objective is to evaluate whether the proposed CSHTR can cope with severe cold-start

situations. For the first objective, as the goal is to evaluate the recommenders’ ability to

recommend user preferred items, therefore, only is used to evaluate the resulting

recommendation list . For the second objective, because in the cold-start situations

recommender systems usually don’t possess sufficient rating data and | | might be very

small, therefore, the evaluation standard for recommenders in severe cold-start situations

is relaxed so that will be used to evaluate .

Page 92

In order to evaluate the performances of different recommenders based on and

, recommendation list based evaluation metrics (i.e. classification accuracy metrics)

such as precision and recall, Breese Score, Half-life, and etc. (Herlocker et al., 2004,

Schein et al., 2002) can be utilised, for more details about these metrics please refer to

Section 2.4.1.2. In this thesis, the precision and recall metrics are used for the evaluation,

and their formulas are listed below:

| |

| | 

(3.19)

| || |

(3.20)

Note, for the evaluation of CSHTR (for cold-start problems), is replaced with

for both precision and recall measures.

In order to provide a general overview of the overall performances, 1 metric is

used to combine the results of Precision and Recall, details about the 1 metric is

provided in Section 2.4.1.2.:

12

(3.21)

For the computation efficiency evaluation, the average time required by

recommenders to make a recommendation will be compared.

Page 93

3.3.3.4 Experimental Results

Corresponding to the two evaluation objectives addressed in Section 3.3.3.3, two

different testing datasets are constructed for different evaluation objectives. Each record

in the datasets consists of the testing ratings of one user . The first testing dataset

(denoted as NOR_testing) is constructed by randomly choosing 10,000 users from the

278,858 users in the entire Book-Crossing dataset mentioned in Section 3.3.1. The first

dataset is used to evaluate recommenders in normal situations (i.e. without specific cold-

start problems), where the neighbourhoods of these users in the dataset NOR_testing can

be formed (or found) with high item preference similarity. The second testing dataset

(denoted as CS_testing) is used to evaluate recommenders in severe cold-start situations.

This set is constructed by choosing 2,000 users with item preferences dissimilar to all

user clusters (i.e. unable find to form neighbourhoods with similar item preferences, i.e.

c ). The details of the two testing datasets are given in Table 3.2.

Table 3.2. Information for the two different testing datasets

NOR_testing CS_testing

Number of users 10000 2000

Average number of explicit ratings

9.77 3.24

Average number of implicit ratings

18.45 8.47

Experiment Results for dataset NOR-testing

We start by evaluating the recommenders’ recommendation qualities for the

normal user set. We let each recommenders recommend a list of items to each of these

10,000 users, and different values for ranging from 5 to 25 are tested. For this part of

Page 94

the experiment, CSHTR is excluded as it is designed specifically to be operated in cold-

start situations.

The results of this part of the experiment are shown in Figure 3.6, Figure 3.7 and

Figure 3.8. It can be observed from the figures that, for all the three evaluation metrics,

the proposed HTR technique achieves the best result among all the recommenders. In the

case of using only explicit rating data, the recommendation quality of HTR (i.e. HTR_E)

still outperforms other recommenders even slightly degrading compared with using both

explicit and implicit rating data (i.e. HTR performs the best and HTR_E performs the

second best).

The standard item-based CF recommender (IR) performed very similarly to the

Slope One recommender (SO), however, it seems that slope one recommender is slightly

better in recommending longer item lists.

In the experiment, the clustering-based CF recommender (IRC) performed better

than the standard one (IR). The only difference between these two recommenders is in

the candidate item list formation process. The standard item-based CF uses all items

from the dataset as its candidate item list (i.e. \ ) , whereas the clustering-based

version uses only items within a user cluster (i.e. \ ). Intuitively,

the clustering-based CF might perform worse than the standard one, because its

candidate item list is formed from a cluster that is only a subset of the entire item set,

some potential promising items might be excluded and, thus, will not be recommended.

However, based on our observation, many of these excluded items are noises generated

from the item similarity measure (some item similarity measures might generate

prediction noise, please refer to (Deshpande and Karypis, 2004) for more information),

therefore, removing these items from the candidate list can actually improve the

recommendation quality. The proposed HTR also gets benefits from the clustering

Page 95

strategy as it generates recommendations from the candidate item list formed from a

cluster.

Figure 3.6. Recommender evaluation with precision metric

0

0.02

0.04

0.06

0.08

0.1

0.12

0.14

0.16

0.18

0.2

#5 #10 #15 #20 #25

Precision

top k recommened items

IRC IR SO HTR

TPR HTR_E ITR

Page 96

Figure 3.7. Recommender evaluation with recall metric

Figure 3.8. Recommender evaluation with F1 metric

0

0.02

0.04

0.06

0.08

0.1

0.12

#5 #10 #15 #20 #25

Recall

top k recommened items

IRC IR SO HTR

TPR HTR_E ITR

0

0.02

0.04

0.06

0.08

0.1

0.12

#5 #10 #15 #20 #25

F1

top k recommened items

IRC IR SO HTR

TPR HTR_E ITR

Page 97

Figure 3.9. Computation efficiency results for different recommenders (average

seconds per recommendation)

From the results shown in the Figure 3.6, Figure 3.7 and Figure 3.8, we can see

that the TPR technique described in (Ziegler et al., 2004) performed the worst among all

recommenders in our experiments. This is because TPR uses only implicit ratings as its

data source and generates recommendations only based on taxonomy preferences,

whereas in our evaluation scheme (see Section 3.3.3.3) recommendations are evaluated

based on these explicitly preferred items (i.e. ). In order to make the proposed HTR

and Ziegler’s TPR more comparable, we modified TPR by adding the item-based CF

component into TPR resulting in the new recommender ITR. ITR performed better than

the standard TPR as it included the item preference in its recommendation making

process. However, it is still worse than all other recommenders (i.e., TPR performs the

worst and ITR performs the second worst). The difference between HTR and ITR is that

the method to compute the taxonomy preferences is different (they use the same method

to compute the item preferences). The result of HTR outperforming ITR indicates that

0.0017 

5.6664 

5.0825 

0.0861 

1.5441 

0.0473 

2.0355 

0

1

2

3

4

5

6

IRC IR SO HTR TPR HTR_E ITR

seconds

recommender types

Page 98

users’ item preference is also helpful for generating users’ taxonomy preference. The

proposed HTR technique considers the item preference implication when generating the

taxonomic preferences (i.e. the taxonomic preferences are extracted from user clusters

that are divided based on users’ item preferences). In contrast, TPR generates users’

taxonomic preferences purely from taxonomy data without using any of the users’ item

preferences.

In the experiment, the recommender with the best computation efficiency is the

clustering based CF (IRC) as shown in Figure 3.9. Computation efficiency results for

different recommenders (average seconds per recommendation), and it is much faster

than the standard CF because its candidate item list is much smaller. The proposed HTR

methods (HTR and HTR_E) perform the second and the third best, as they spent a bit

more time on predicting taxonomic preferences comparing to IRC. However, this extra

computation complexity is trivial, because most of these computations (i.e. computing

_ for each user cluster) can be done offline. HTR_E performed slightly better

than HTR because it uses less data (only explicit ratings) to make recommendations.

Ziegler’s TPR is computation expensive because it needs to convert all users and items

into high dimensional taxonomy vectors. ITR performed slightly worse than TPR

because it needs to compute extra item preference predictions using standard CF

technique. Standard item-based CF (IR) technique is the most inefficient one among all

the recommenders, as it needs to build entire candidate item list from scratch and

compute correlations between the user profile and the candidate items for making each

recommendation. In contrast, Slop One recommender (SO) offers a slight advantage in

computation efficiency by pre-computing the correlations between the user profiles and

the items in advance, however, as forming the candidate item lists is still a lengthy

Page 99

process, it is still not as efficient as other techniques with pre-computed candidate item

lists (extracted from the pre-computed user clusters).

Parameterisation Analysis for HTR

In Section 3.3.2, the suggested assumption for the relation between item

preference and item taxonomic preference is verified, and the relation is applied and

utilised by the proposed HTR and CSHTR techniques. In the last section, our experiment

demonstrated that the proposed HTR technique is superior to other existing

recommenders in both recommendation quality and computation efficiency. However,

what has not been shown in the experiment is whether HTR’s superior recommendation

quality in the experiment is indeed resulted from the integration of item preference and

item taxonomic preference.

In order to demonstrate the integration of item preference and item taxonomic

preference does affect the recommendation quality of HTR, we evaluate HTR’s

performance with different settings to the adjustment parameter (in Equation (3.11)).

As described in Section 3.2.5, the adjustment parameter controls the weight

distribution between item preference and item taxonomic preference in the final ranking

score computation. When the value of equals to 1, HTR considers only item

preference (i.e. , ) in its recommendation making process, and therefore behaves

similar to the standard Item-based CF (i.e. IR or IRC). Conversely, when the value of

approaches to 0, only item taxonomic preference (i.e. , ) is considered in the final

ranking computation.

Besides , the adjustment parameter in Equation (3.7) is also investigated in

this section for its relation to the recommendation quality of HTR. It is mentioned in

Section 3.2.4.3 that is used to adjust the weights of personal level taxonomic

Page 100

preference (i.e. _ , ) and cluster level taxonomic preference (i.e.

_ , ) in the final taxonomic preference score computation (i.e. , ),

therefore by experimenting different values in the performance evaluation we can

investigate whether the integration of personal and cluster level taxonomic preference is

beneficial to HTR.

Figure 3.10. F1 results for HTR with different and configurations.

In the experiment, different value combinations of and have been used to

configure HTR, and the performance results (captured and measured based on the F1

metric) obtained from different HTR configurations are depicted in Figure 3.10.

In order to ensure the fairness of the investigation, for different value

combinations of and , all other parameter configurations are kept the same (see

Section 3.3.3.2 for the detail of HTR’s parameter configurations). In this experiment, we

let HTR to recommend top 10 items (i.e. 10) as this setting has resulted in the best

00.2

0.40.6

0.81

0.03

0.04

0.05

0.06

0.07

0.08

0.09

0.1

00.2

0.40.6

0.81

α1

F1

0.03‐0.04 0.04‐0.05 0.05‐0.060.06‐0.07 0.07‐0.08 0.08‐0.09

Page 101

recommendation quality in previous experiments (i.e. in comparison to

5, 15, 20 or 25).

Figure 3.11. F1 results for HTR with different configurations ( 0.2)

Figure 3.12. F1 results for HTR with different configurations ( 0.8)

0

0.01

0.02

0.03

0.04

0.05

0.06

0.07

0.08

0.09

0.1

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

F1   

0.08

0.082

0.084

0.086

0.088

0.09

0.092

0.094

0.096

0.098

0.1

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

F1   

α1

Page 102

In order to provide better observations for ’s effect on the performance of HTR,

a 2D graph showing HTR’s performance with different configurations of ( 0.2

remains static) is extracted from Figure 3.10 and depicted in Figure 3.11. It can be

observed from Figure 3.10 and Figure 3.11 that, HTR yielded very low recommendation

quality when  approaches to 0 or 1, and it performed well when is between 0.2 and

0.8 (despite the value of ). Hence, the result suggests that the integration of item

preference and item taxonomic preference information is indeed significantly beneficial

for making quality recommendations.

Similar to Figure 3.11, Figure 3.12 depicts a 2D graph showing HTR’s

performance with different configurations of ( 0.8 remains static). It can be

observed from Figure 3.10 and Figure 3.12, HTR achieved best performance in

recommendation quality when both personal level taxonomic preference (i.e.

_ , ) and cluster level taxonomic preference (i.e. _ , )

are considered (i.e. 0.2). By comparing Figure 3.11 and Figure 3.12, it can be seen

that has less effects to HTR’s performance than , that is, the range of performance

difference for is about 0.0045 (from 0.0905 to 0.0959) and for is about 0.06 (from

0.035 to 0.0959). Even though integrating both cluster and personal level taxonomic

preference resulted better recommendation quality, however, the amount of

improvement achieved is small (i.e. 0.0045). Therefore, if further computation efficiency

optimisation is required, little recommendation quality advantage can be sacrificed by

skipping the computation of personal level taxonomic preference (i.e. set 1). The

computation of cluster level taxonomic preference is more efficient than personal level

taxonomic preference as it can be pre-computed offline and shared by multiple users,

Page 103

hence, if only cluster level taxonomic preferences are required, the efficiency of the

proposed HTR technique can be effectively improved.

Based on the experiment results shown in Figure 3.10, Figure 3.11 and Figure

3.12, it can be concluded that a recommender’s recommendation quality can be

improved and benefited by:

Integrating users’ item taxonomic preferences and item preferences together

into recommendation making.

Integrating cluster level and personal level taxonomic preferences together

for extracting users’ item taxonomic preferences.

Experiment Result for dataset CS-testing

In this part of experiment, we evaluate the performance of the proposed CSHTR

under cold-start conditions (i.e. with the user set CS-testing). The baseline recommender

for this evaluation is Ziegler’s TPR, because TPR is the only technique among the others

that is specifically designed for making recommendations in severe cold-start situations.

Except for the CSHTR and TPR, all the other six recommenders listed in Section 3.3.3.1

are not included in this evaluation as they are generally sensitive to cold-start problems

and do not perform well in the cold-start situation (it is because they make

recommendations mainly based on explicit item rating data).

The evaluation results for the cold-start situation are shown in Figure 3.13,

Figure 3.14 and Figure 3.15. The comparison of computation efficiency between

CSHTR and TPR is shown in Figure 3.16. It can be seen from the results that, the

recommendation quality of the proposed CSHTR is better than that of Ziegler’s TPR. It

suggests that the use of common item preferences in the target users’ belonging cluster is

beneficial for alleviating cold-start problems. Moreover, CSHTR offers much better

Page 104

computation efficiency than does TPR, mainly because CSHTR uses the expensive

taxonomy vector similarity computation only for computing the similarity between the

target user and the candidate items, whereas TPR computes the similarities for all users

within their neighbourhood as well as the candidate items.

Figure 3.13. Recommender evaluation under cold-start situations with precision

metrics

Figure 3.14. Recommender evaluation under cold-start situations with recall metrics

0

0.001

0.002

0.003

0.004

0.005

0.006

0.007

0.008

0.009

#5 #10 #15 #20 #25

Precision

top k recommened items

CSHTR

TPR

0

0.001

0.002

0.003

0.004

0.005

0.006

0.007

0.008

0.009

#5 #10 #15 #20 #25

Recall

top k recommened items

CSHTR

TPR

Page 105

Figure 3.15. Recommender evaluation under cold-start situations with F1 metrics

Figure 3.16. Computation efficiencies for CSHTR and TPR

3.4 CHAPTER SUMMARY

In this chapter, we investigated the implicit relations between users’ item

preferences and taxonomic preferences, suggested and verified that users that share

0

0.001

0.002

0.003

0.004

0.005

0.006

0.007

#5 #10 #15 #20 #25

F1

top k recommened items

CSHTR TPR

0.0426 

0.4779 

0

0.1

0.2

0.3

0.4

0.5

0.6

CSHTR TPR

seconds

recommender types

Page 106

similar item preferences may also share similar taxonomic preferences. Based on this

investigation, we proposed a novel, hybrid technique HTR to automated

recommendation making based upon large-scale item taxonomies that are readily

available for diverse ecommerce domains today. An HTR’s extension, CSHTR, is also

proposed specifically for alleviating the cold-start problems.

HTR and CSHTR produce quality recommendations by incorporating both users’

taxonomic preferences and item preferences. Moreover, these two proposed techniques

can utilise both explicit and implicit ratings for recommendation making, and hence they

are less prone to suffer from the cold-start problems. We have compared the proposed

HTR technique with some standard benchmark techniques, such as item-based

recommender and some advanced modern techniques such as TPR. We have conducted

extensive experiments that demonstrated that the proposed HTR outperforms other

recommenders in both recommendation quality and computation efficiency. In addition,

our evaluation has shown that the proposed CSHTR method performs effectively under

the cold-start situations, and it outperformed the baseline technique, TPR, in both

recommendation quality and computation efficiency.

Page 107

Chapter 4 

4Distributed Recommendation Making

In Chapter 3, the possibility of alleviating the cold-start problem by enriching

information resources with additional data facets is examined and demonstrated.

Specifically, the widely available data source, item taxonomy, is investigated and studied

for its applicability in recommender systems. We identified an implicit relation between

users’ item preferences and item taxonomic preferences, and successfully utilised this

relation to alleviate the cold-start problem as well as improve recommendation quality.

In this chapter, another strategy for alleviating the cold-start problem is explored,

that is, to increase data volume of recommenders via allowing them to share and

exchange data and resources with each other over a distributed environment. As

mentioned previously, most of the existing recommender systems are implemented for

one organisation (i.e. business to customer (B2C) recommenders), and in general one

single organisation may not possess sufficient information or data for analysis in order to

give their customers precise and high-quality recommendations (hence results in the

cold-start problem). Therefore, it can be beneficial if organisations can share their

resources (i.e. products and customer database) and recommendations boundlessly (i.e.

build recommendation systems at inter-organisational level), and more importantly, great

business value might be generated by the resource sharing among the organisations.

In this chapter, we present a framework for distributed information sharing

among recommenders. The proposed distributed framework is different from existing

distributed recommender systems. The existing distributed recommender systems are

mainly designed for C2C (Customer to Customer) based applications (such as P2P and

Page 108

file sharing applications), the proposed distributed recommender system introduces

additional B2B (Business to Business) features on top of the standard B2C (Business to

Customer) recommender systems.

This chapter roughly consists of two parts. In the first part, we explain the

rationale of the proposed distributed recommender system, and then describe and discuss

the system models and infrastructures. In the second part, we describe and discuss in

details a recommender peer profiling and selection strategy designed for the proposed

distributed recommender system.

4.1 RELATED WORK

Section 2.3 has comprehensively reviewed existing distributed recommender

systems. This section mainly discusses and compares works that bear strong

resemblance to the proposed distributed recommender framework.

Wei (2003) has proposed an multi-agent based recommender system in which

the recommender system is considered as a marketplace consisting of one auctioneer

agent and multiple bidder agents. Each bidder agent is considered as a recommendation

algorithm that is capable of generating recommendations independently, and within the

marketplace these bidder agents compete to each other for short-listing their

recommendations. The task of auctioneer agent is to incorporate the bids of the bidder

agents and generating the most suitable result to the users. Essentially, Wei’s approach is

a hybridised recommender system designed based on the concept of multi-agent system.

Even though Wei’s system is designed to work within a single organisation and is not

considered as a distributed recommender system, it is still mentioned here as it takes the

concept of decentralised decision making into consideration (i.e. making

recommendations based on the cooperation of multiple recommender agents).

Page 109

As mentioned previously, most existing studies on distributed recommender

systems are mainly designed for peer-to-peer (P2P) or file sharing applications (which

usually adhere to C2C paradigm). We have discussed many relevant studies in this

category in Section 2.3, here we would like to address Awerbuch’s (2005) work in

particular, as it provides a generalised view to these distributed recommenders.

Awerbuch suggested a formalised model for the C2C distributed recommender systems.

In Awerbuch’s model, for the distribute system with users and items, there will be

recommender systems (i.e. agents or peers), and each of the recommender agents will

associate with exactly one user. Each recommender works on behalf of the associated

user either to trade recommendations with other agents or probe the items on its own.

Each recommender aims to finally discover the items preferred by the associated user,

where . In Awerbuch’s opinion, from the perspective of the entire distributed

recommender system, the goal is rather similar to the ‘matrix reconstruction’ proposed

by Drineas et al. (2002); the overall task is to reconstruct an user preference

matrix in a distributed fashion. It can be observed that many distributed recommender

systems can fit into such model.

Generally, the goal of these C2C based distributed recommenders is to avoid

central server failure and protect user privacy (no central database containing

information about customers) (Awerbuch et al., 2005, Castagnos and Boyer, 2007, Han

et al., 2004, Liu et al., 2007, Sorge, 2007, Tveit, 2007, Vidal, 2004, Wang et al., 2006,

Ziegler and Golbeck, 2007) . However, most of them are not aimed at improving their

effectiveness or the recommendation quality. In contrast, the goal of the distributed

recommender system we proposed is aiming at improving the recommendation quality

and alleviating the cold-start problem. Hence, the infrastructure of the proposed

distributed recommender system is different from Awerbuch’s model as well as many

Page 110

other existing systems. Our system contains a set of classical recommenders, and each of

them serves their own set of users. Our goal is to improve the recommendation quality of

these recommenders by allowing them making recommendations for others in a

decentralised fashion. Thus, for the profiling and selection problem, we proposed a more

sophisticated strategy rather than random sampling for recommender peers to explore

others.

Moreover, recommender systems and information retrieval (IR) systems are

generally considered similar research fields (Herlocker et al., 2004, Sarwar et al., 2002),

since both of them try to satisfy users’ information needs by either retrieving the most

relevant documents or recommending the most preferred items to users. Information

retrieval retrieves documents based on users’ explicit queries, while recommender

systems recommend items or products based on users’ previous behaviour. In distributed

IR (Christoph, 1997, Kretser et al., 1998), the entire document collection is partitioned

into sub-collections that are allocated to various provider sites, and the retrieval task then

involves:

Querying minimal number of sub-collections (to improve the efficiency),

and ensure the selected sub-collections are significant to uphold the retrieval

effectiveness.

Merging the queried results (fusion problem) that incorporate the differences

among the sub-collections in such a way that no decrease in retrieval

effectiveness is effectuated with respected to a comparable non-distributed

setting.

For distributed recommender systems, the recommender peer selection and

recommendation result merging are also two important tasks. In fact, one of the major

research focuses of our research is to design an effective recommender peer profiling and

Page 111

selection strategy. The selection criteria for distributed IR including the efficiency

(selecting minimal number of sub-collections) and effectiveness (retrieving the most

relevant documents) is similar to the criteria for the proposed distributed recommender

system. However, in distributed IR, the collection selection is content-based (Christoph,

1997, French et al., 1999, Kretser et al., 1998) and it requires the sub-collections provide

or use sampling techniques to get sub-collection index information (e.g. the most

common terms or vocabularies in the collection) and statistical information (e.g.

document frequencies). In contrast, the proposed selection technique requires no content

related information about recommender peers (assuming recommender peers share

minimal knowledge to each other), the proposed selection algorithm is based on the

observed previous performance (i.e. how well a recommender peer’s recommendations

satisfy the users) about each of the recommender peers.

4.2 ECOMMERCE-ORIENTED DISTRIBUTED RECOMMENDER

As mentioned earlier, the goal of the proposed distributed recommender system

is to allow standard recommenders to overcome cold-start problem and improve

recommendation quality by cooperating, interacting and communicating with

recommenders of other parties (e.g. other ecommerce sites). Hence, the proposed system

is designed to contain of a set of recommenders from different sites and each of these

recommenders is associated with their own users. It is important to note that is possible

that a user might visit multiple sites, and therefore two or more recommenders may share

common users. Similar to the centralised paradigm, each recommender peer in the

proposed system still serve its own users in a centralised fashion (i.e. the recommender

stores all its user and product data in a central place within the recommender). However,

in the proposed system, the recommender peers can enrich their information resources

Page 112

by communicating and cooperating with each other. A general overview of the proposed

system is depicted in Figure 4.3.

Since the proposed distributed recommender system is designed to benefit

ecommerce sites (rather than focusing on helping users to gain more controls on

recommenders), we therefore named our system as ‘Ecommerce-oriented Distributed

Recommender System’, and abbreviated it to EDRS. We also abbreviate the standard

Distributed Recommender System to DRS and Centralised Recommender System to

CRS in order to clarify and differentiate the three different system paradigms.

Before explaining the proposed distributed recommender framework in more

detail, some general differences among the EDRS, DRS and CRS are investigated. In

particular, these systems are compared according to the following aspects:

Ecommerce Model: Based on the general ecommerce activities and

transactions involved in the recommenders’ host application domains, we

can roughly categorise them into three different models, namely, Business-

to-Business (B2B), Business to Customer (B2C) and Customer to Customer

(C2C). In B2B model, activities (e.g. transactions, communications and

interactions) mainly occur among businesses. In the B2C model, activities

are mainly between businesses and customers, and the most typical example

is activities of E-businesses serving end customers with products and/or

services. Finally, the C2C model involves the electronically facilitated

transactions between consumers. A typical example is the online auction

(e.g. eBay), in which a consumer posts an item for sale and other consumers

bid to purchase it.

Architectural Style: An architectural style describes a system’s layout,

structure, and the communication of the major comprising system modules

Page 113

(or software components). Over past decades, many architectural styles have

been proposed, such as, Client-Server, Peer-to-Peer (P2P), Pipe and Filter,

Plugin, Service-oriented, etc. Client-Server and Peer-to-Peer are the two

major architectural styles related to our thesis, and therefore will be

explained in more details. The Client-Server architecture usually consists of

a set of client systems and one central server system, client systems make

service requests over a computer network (e.g. internet) to the server system,

and the server system fulfils these requests. Peer-to-Peer architecture

consists of a set of peer systems interacting with each other over a computer

network, and it does not have the notion of clients and servers, instead, all

peer systems operate simultaneously as both servers and clients to each other.

Communication Paradigm: Based on how two types of entities

communicate with each other within a system, three major communication

paradigms have been proposed, and they are One-to-One, One-to-Many and

Many-to-Many communication paradigms (or relationships). In One-to-One

communication paradigm, communication occurs only between two

individual entities, example applications include: e-mail, FTP, Telnet, etc. In

contrast, a website that displays information accessible by many users is

considered having a One-to-Many relationship. In Many-to-Many paradigm,

entities communicate freely with many others, example applications include:

file sharing (multiple users to multiple users), Wiki (multiple authors to

multiple readers), Blogs, Tagging, etc.

Figure 4.1 shows a general overview of a standard centralised recommender

system (i.e. CRS). The host application of CRS is usually an ecommerce site (e.g.

Amazon.com, Netflix.com, etc.), which possesses all user/product relevant information,

Page 114

and the recommender then utilises all the information from the site to make personalised

recommendations to the site’s users and further create business values to the ecommerce

site. As the nature of the CRS is to serve the users (i.e. customers) and to satisfy the users’

information needs to the ecommerce site (i.e. business), it can be considered as adhering

to the B2C paradigm. It is usually implemented based on the Client-Server architecture

because the entire recommendation generation process occurs only within the central

server, and users interact with the recommender though thin clients (e.g. web browsers)

whose major functions are presenting users the recommendations generated from the

server and sending users’ information requests to the server. In the most common case,

all users of a site are served by a single recommender, therefore, the communication

paradigm between recommenders and users in CRS is considered as One-to-Many.

Figure 4.1. Classical centralised recommender system

Page 115

The standard distributed recommender system (DRS), as depicted in Figure 4.2,

differs from CRS in all the three of the mentioned aspects. First of all, it emphasises

users’ privacy protection by preventing personal user data being gathered and used (or

misused) by ecommerce site owners (or businesses), hence adheres to the Customer-to-

Customer model (as Business entities are evicted from the system for privacy protection).

It is shown in Figure 4.2 that, a standard distributed recommender system associates

every user in the system with a recommender peer serving the user’s personal

information needs, hence the relationship between the user and recommender is

considered as One-to-One. On the other hand, in order to make better recommendations

to its user, a recommender peer might need to communicate with other peers to exchange

its user’s data (in a privacy protected way) with other peers or to get recommendations

from other peers because there is no central place for storing all users’ data. The

relationship among recommender peers in the DRS is considered as Many-to-Many, as a

peer can both communicate to and be communicated by many other peers. Finally,

because all recommender peers are equipped with similar set of functionalities (i.e.

gather information from others and making recommendation to its user) and operate

independently and autonomously from others, therefore, they are commonly modelled

and implemented using Peer-to-Peer architectural style.

Page 116

Figure 4.2. Standard distributed recommender system

The proposed Ecommerce-oriented Distribute Recommender System (EDRS)

(depicted in Figure 4.3), can be thought of as a combination of the two systems

(centralised recommender and DRS) described above. Similar to the DRS, EDRS

consists of a set of recommender peers and a set of users. However, while one user is

associated with exactly one recommender peer in the standard distributed recommender

system, the proposed system can be considered as a set of centralised recommender

systems cooperate together to serve their own set of users, and therefore each

recommender peer needs to interact (i.e. make recommendations to) with multiple users.

Moreover, it is also possible that in our system a user is associated with more than one

recommenders (i.e. he or she can visit multiple sites); for instance, a book reader might

try to find a book in both Amazon.com and Book.com. As a recommender peer in our

Page 117

system can serve multiple users and a user can make recommendation requests to

multiple recommender peers, the relationship between users and recommender peers is

considered as Many-to-Many. As mentioned previously, the recommender peers in

EDRS might interact and cooperate with each other to improve their recommendation

quality, and hence, apart from the Many-to-Many relationship between users and

recommender peers, another Many-to-Many communication relationship exists among

the peers.

Since EDRS is still designed for normal ecommerce sites, such as e-book stores

like Amazon.com, its major ecommerce model is therefore the same as CRS, that is,

Business-to-Customer. Besides, since EDRS introduces additional communication and

cooperation for recommenders of different sites, it is expected that the cooperation of

these recommenders (also their sites) will confirm to the Business-to-Business based

model.

The implementation of the proposed EDRS involves both Peer-to-Peer and

Client-Server architectural styles. Client-Server architecture is employed to model a

recommender peer (i.e. the server) and its users (i.e. the clients). Similar to the

centralised recommender, the entire recommendation generation process is done by the

recommender situated at the server side, and the users make requests to the

recommender through thin clients such as web browsers. The architectural style for the

network among the recommender peers is modelled with Peer-to-Peer architecture. As

mentioned previously, Peer-to-Peer based architecture assumes that the peers are

independent and autonomous from each other, and especially they should be loosely

coupled. Such a definition is suitable for modelling the relationship between the

recommender peers’ host sites, as they are both logically and physically independent and

autonomous from each other (as they are different ecommerce sites and organisations).

Page 118

While both DRS and the proposed EDRS can be modelled with the Peer-to-Peer

architecture, the recommender peers in EDRS are more strongly coupled together than in

standard DRS. This is because the recommender peers in EDRS need to

gather/distributed information and suggestions from/to each other in a timely and

effective fashion to achieve their common goal (i.e. satisfy their users’ information and

recommendation needs).

To the best of our knowledge, the concept of the proposed EDRS has not yet

been mentioned and investigated by other studies. In addition, it is different from

existing recommender systems (both centralised and distributed ones) at several high

level aspects. Table 4.1 summarises these differences.

Table 4.1. High level aspect differences among recommender system paradigms

Ecommerce Model

Architectural Style Communication Paradigm

CRS B2C Client-Server One-to-Many (recommender to user)

DRS C2C Peer-to-Peer

Many-to-Many (recommender to recommender) One-to-One ( recommender to user)

EDRS B2C, B2B Client-Server, Peer-to-Peer

Many-to-Many (recommender to recommender/

Recommender to user)

Page 119

UserMany to ManyRelationship

Many to Many Relationship

Recommender Agent (reside at Ecommerce site’s server)

BusinessBusiness to

Customer

Business

to

Figure 4.3. Proposed distributed recommender system

4.2.1 General Interaction Protocol

As mentioned earlier, the interaction, communication and cooperation of the

recommender peers in the proposed EDRS can be modelled with the Peer-to-Peer based

architectural style. In particular, the ‘Contract Net Protocol’ (CNP) is employed as the

foundation for modelling the system, which provides the basis for coordinating the

interaction and communication among the recommender peers. Contract Net Protocol is

a high level communication protocol and system modelling strategy for Peer-to-Peer

Page 120

architectural based systems (or other distributed systems) (Smith, 1981, Weiss, 1999). In

CNP, peers in the distributed system are modelled as nodes and the collection of these

nodes is referred to as a contract net. In CNP based systems, the execution of a task is

dealt with as a contract between two nodes, each node plays a different role, one of them

is the manager role and the other is the contractor role. The role of a manager is

responsible for monitoring the execution of a task and processing the results of its

execution. On the other hand, the role of a contractor is responsible for the actual

execution of the task. It is important to note that the nodes are not designated a priori as

contractors or managers, rather, any nodes may take on either roles dynamically based

on the context of their interaction and task execution (Weiss, 1999, Smith, 1981). A

contract is established by a process of mutual selection based on a two-way transfer of

information. In general, available contractors evaluate task announcements made by

managers and submit bids on those for which they are suited. The managers evaluate the

bids and award contracts to the nodes (i.e. contractors) that they determine to be most

qualified (Smith, 1981).

In the case of the proposed EDRS, the recommender peers are modelled as the

nodes in the contract net. Depending on difference circumstances, each recommender

peer plays manager role and contractor role interchangeably. When a recommender peer

makes requests for recommendations to other peers, it is considered as a manager peer.

Conversely, the recommender peer that receives a request for recommendations and

provides recommendations to other peers is considered as a contractor peer. The roles of

the manager peer and the contractor peer and their interactions are depicted in Figure 4.4.

Page 121

Figure 4.4. High level interaction overview for EDRS (based on contract net protocol)

The communication steps involved in the interaction are indicated by the

numbers in Figure 4.4 and explained as follows:

(1) User sends a request for recommendations. The recommender peer who

received the request and is responsible for making the recommendation to

the user is considered to be in the manager role.

(2) Based on the user’s request and profile, the manager peer selects suitable

peer recommenders to help it on making better recommendations to the user.

(3) The manager peer makes requests to the peers for recommendation

suggestions. The request message may only contain the user’s item

preferences (i.e. the user’s rating data); however, the identity of the user is

anonymous for privacy protection.

(4) Each contractor peer generates recommendations based on the received

request.

Page 122

(5) The contractor peers send back their recommendation suggestions to the

manager peer.

(6) After the manager peer received the suggestions from the contractors, it then

synthesises and merges these recommendation suggestions.

(7) Based on the synthesised recommendation suggestions from the contractor

peers (might also include the manager peer’s own recommendations) the

manager peer generates the item recommendations to the user.

(8) When the user received the recommendations, he or she might supply

implicit or explicit ratings to the recommendations. That is, the user might

provide indications about whether he or she likes or dislikes one or more

items in the recommendation list.

(9) Based on the user rating feedbacks, the manager peer can objectively

evaluates each of the peers’ (i.e. contractors’) performances to the

recommendation suggestions they supplied and update its profiles about

these peers.

(10) The manager peer sends feedbacks and rewards to the contractor peers based

on their performances to the task.

(11) When the contractor peers received feedbacks about the performances of

their recommendation suggestions, they then update their profiles about the

manage peer in order to improve their future suggestions.

From Figure 4.4, it can be seen that when a recommender peer is requested to

make recommendations for a user, it acts as a manager peer. In the role of a manager

peer, the recommender first generates a strategy about how and what to recommend to

the user based on the user’s profile and request. Then the recommender chooses a set of

recommender peers (in this context, they act as contractor peer) based on the profiles of

Page 123

peer recommenders, and finally makes requests for recommendations to these selected

contractor peers. When these selected contractor peers received the requests, they then

construct and return their recommendation suggestions based on the requests received

and the manager peer’s profile (e.g. preferences, domain of interests, and

trustworthiness). After the manager peer received the recommendations returned from

the contractor peers, it then merges the recommendations (also include recommendations

from itself) and return to the user. According to the recommendations received from the

manager peer, the user might either explicitly or implicitly give feedbacks or ratings

about the recommendations to the manager peer. After receiving the user’s feedback, the

manager peer will evaluate the performance of each of the selected contractor peers,

update its profiles about them, and then construct the feedbacks and make rewards to the

contractor peers. Finally, the contractor peers will update theirs profile about the

manager peer based on the given rewards and feedbacks.

In order to carry out the proposed interaction described above, the following

tasks need to be considered.

Recommender Peer Selection: After a manager peer received a request

from a user, it needs to determine a subset of recommender peers from all

available recommender peers to consult for recommendations. A mechanism

is required so that:

(1) The number of peers selected is minimised (to ensure efficiency); and

(2) The user’s satisfaction to the collected recommendations is maximised.

Recommendation generation: As the system is loosely coupled (as these

recommender peers are from different ecommerce sites), therefore, each

peer does not hold detailed knowledge about the data collections, the

operations and functions of other peers. Hence, depending only on the

Page 124

request sent from the manager peer is not sufficient for the contractor peers

to generate quality recommendations. Therefore, the contractor peers

generate recommendations based on both the content of the request and their

profiles about the manager peer as well.

Recommendation merge: Mechanisms are required for the manager peer

to synthesise the recommendation gathered from different contractor peers,

such that:

(1) The synthesised recommendation should result in a high level

satisfaction from the target user. In general, the synthesised

recommendation should have better quality than the recommendation

generated by the manager peer itself; and

(2) The contributions of the contractor peers to the synthesised

recommendation need to be balanced (i.e. without degrading the quality

of the recommendation, the final recommendation should be constructed

by considering as many contractors’ recommendations as possible), so

that the manager peer can extend or update its knowledge to as many

peers as possible based on the user feedbacks to the recommendation.

Peer feedback and profile update: The major source that recommenders

can learn about each other is from the user feedbacks. Based on the user

feedbacks to a particular recommendation, the manager peer needs to

evaluate the performances of each individual contractor peer, and further

acquire better understanding about them. Moreover, the manager peer also

needs to supply feedbacks and rewards to the contractor peers, so that the

contractor peers can learn the manager peer’s preferences as well.

Page 125

Hence, it can be observed that each of the recommender peers in the EDRS

need to maintain two set of peer profiles. The first set of profiles is the

contractor peer profile set which is used when the peer is in the manager role

and other peers are in contractor role. In contrast, the second set of profiles is

the manager peer profile set which is used when the peer is in the contractor

role and other peers are in manager role. However, due to the limited scope

of this thesis, only the contractor peer profiles are considered in our thesis

and experiments. Thus, we allow a recommender peer in the manager role to

select other contractor peers based on its contractor peer profile set, and as

the contractor peers maintain no profiles about the manager peer, it is

assumed that they will generate recommendations to the manager peer only

based on the manager peer’s current request/query and not past behaviours.

Therefore, given two different manager peers with same requests, a

contractor peer will generate same recommendations to them.

Among these four proposed tasks mentioned above, recommender peer

profiling and selection for manager peers is the major focus of this thesis, and a novel

contractor peer profiling and selection strategy is proposed, discussed and investigated in

Section 4.3. Section 4.4 describes a simple technique for a manager peer to merge

recommendations generated from multiple contractor peers to form a single

recommendation to the target users. Due to the limited thesis scope, the strategy required

for contractor peers to profile manager peers are not included in this thesis.

4.3 PEER PROFILING AND SELECTION

Part of the major contributions in this chapter includes a recommender profiling

scheme (for manager peers to profile contractor peers) and a recommender selection

Page 126

algorithm designed for the proposed EDRS. In particular, the recommender peer

selection problem is modelled as the classical exploitation vs. exploration (or k-armed

bandit) problem (Azoulay-Schwartz et al., 2004, John, 1989), in which the recommender

selection for the manager peer has to be balanced between choosing the best known

contractor peers to keep users satisfied and selecting other unfamiliar contractor peers to

obtain knowledge about them. The proposed recommender selection algorithm is based

on evaluating the Gittins Indices (John, 1989) for every recommender peer, and the

indices reflect the average performance, stability and selection frequency of the

recommenders (i.e. contractor peers).

4.3.1 System Formalisation for EDRS

Before explaining the proposed strategies and techniques in detail, a formalised

description of the proposed EDRS is given below.

Similar to the formalisation used in Section 3.2.1, the set of users and items are

denoted by , , … , and , , … , respectively. The proposed

distributed recommender system (EDRS) denoted as Φ contains a set of recommender

peers , , … , , i.e. Φ , , … , . The number of recommender peers is

much smaller than the number of users in our system, i.e. . Each recommender

peer Φ has a set of users denoted as , and a set of items denoted as ,

where

and

Page 127

Moreover, as mentioned previously, some users and items can be owned by more

than one recommender peers such that

and

4.3.2 User Clustering

Intuitively, a large set of users can be separated into a number of clusters based

on the user preferences. Users within the same cluster usually share similar tastes

(Drineas et al., 2002) and a cluster with a large number of users and a high degree of

intra-similarity can better reflect the potential preferences of the users belonging to the

cluster. Thus, a collaborative filtering based recommender can improve its

recommendation quality by searching similar users within clusters rather than the whole

user set (Sarwar et al., 2002, Degemmis et al., 2004). However, different user clusters

often vary in quality. The performance of such clustering based collaborative filtering

system is strongly influenced by the quality of the clusters (Sarwar et al., 2002,

Degemmis et al., 2004). For a given recommender, some users might be able to receive

better recommendations if they belong to a cluster with better quality (the cluster has a

large number of users and a high intra-similarity), whereas some other users may not be

able to get constructive recommendations because the cluster to which they belong is

small and has a low intra-similarity. This situation is closely related to the cold-start

problem (Schein et al., 2002), which occurs when a recommender makes

recommendations based on insufficient data resources. Therefore, even for the same

Page 128

recommender, the recommendation performance might be different for different clusters

of users if different user clusters have different quality. In order to provide good

recommendations to various users, the proposed EDRS allows its recommender peers

(i.e. manager peers) to choose peers (i.e. contractor peers) for recommendations to the

current user based on their performances to a particular user cluster to which the current

user belongs. We expect this design to solve the cold-start problem because a

recommender that is making recommendations to a user who belongs to a weak cluster

can get recommendations from recommender peers who have performed well to that

group of users.

In the proposed EDRS, every recommender peer has its own set of user clusters,

and we denote the set of user clusters owned by Φ as

, , , , … , , , such that , . In addition, for the simplicity of the

system, all user clusters are assumed to be crisp sets, such that , , for

, , ,   , . As different recommender peers have different user sets and

different clustering techniques, the size of their cluster set might vary as well, that is,

, Φ: | | | | (or ).

4.3.3 Recommender Peer Profiling

In this section, we present our approach to profile the recommender peers within

the proposed EDRS. To begin with, the performance evaluation of the recommender

peers is explained. The performance of a recommender peer is measured by the degree

of user satisfactory to the recommendations made by the recommender (Herlocker et al.,

2004, Karypis, 2001, Papagelis and Plexousakis, 2004). In our system, a recommender

peer makes recommendations to a user with a set of items , , , , … , ,

where . Once having received the recommendations, the user then inputs his or

Page 129

her evaluations to each of the items. We use to denote the user’s rating to item

, . The value of is between 1 and 0 which indicates how much the user likes

item , . When closes to 1, it indicates the user highly prefers the item, in contrast

when closes to 0, the user dislikes the item. Hence, each time a recommender peer

generates a recommendation list (e.g. ) to a user, it will get feedback

, … , from the user, where 0,1 . With , we can compute the

recommender peer’s current performance χ to the user by:

∑| |

(4.1)

Equation (4.1) measures the current performance of a recommender peer to a

particular user in the current recommendation round. We can use the average

performance of the recommender to the users in the same cluster to measure its

performance to this group of users. The average performance measures how well the

recommender averagely performed in the past. However, the average performance does

not reflect whether the recommender is generally reliable or not. Hence, we employed

the standard deviation to measure the stability of the recommender. Another factor that

should be taken into account for profiling a recommender is the selection frequency,

which indicates how often the recommender has been selected before. In our system, we

profile each recommender peer from the three aspects: recommendation performance,

stability, and selection frequency. As mentioned previously in this chapter, a

recommender will seek for recommendations from other peers when it receives a request

from a user. Broadcasting the user request to all peers is one solution, but obviously, it is

not a good solution since not all of the peers are able to provide high quality

Page 130

recommendations. In EDRS, the recommender peers (i.e. manager peers) will select the

most suitable peers (i.e. contractor peers) for recommendations based on their profiles.

Therefore, each recommender peers in EDRS keeps profiles to each of the other

recommender peers.

A recommender peer may perform differently to different user clusters.

Therefore, its performances to different user clusters are different. For recommender

Φ which has user clusters, that is, , , , , … , , we use , to

denote the average performance of peer Φ to ’s user cluster , . Hence, we can

use a matrix , to represent the average performance of each of the

other peers to each of ’s user clusters, where |Φ| 1 and | | . is

called as the peer average performance matrix of . Similarly, we use and to

represent the stability and selection frequency of other peers to . , and

, are called as the peer stability matrix and peer selection frequency matrix

respectively. In summary, a recommender ’s peer profile is defined as

, , which consists of the three matrixes representing peer recommender’s

average performance, stability, and selection frequency, respectively.

Initially, the , and of are all zero matrixes, because has no

knowledge about other peers. These matrixes will be updated when a recommender peer

helped (i.e. is in contractor role and is in manager role) to make a

recommendation for a user belonging to (or being classified to) a ’s user cluster

, . Suppose that is the recommendation list returned by . Ideally, is expected

to b a subset of . But usually since and may have different item sets. In

the proposed EDRS, only the items that are in are considered by . Let be the final

recommendation list made by to the user and

Page 131

|  and selected by  be the recommendation list made by and

selected by during the merging process (the major focus of this selection is on peer

profiling, other aspects of the proposed EDRS such as merging recommendations from

different peers will be explained in latter sections). should be a subset of . After the

recommendation is provided to the user, will get a feedback list (i.e. the actual user

ratings to the recommended items) about from the target user. With the user

feedback , Equation (4.1) will be used to compute ’s performance for the

recommendation of this round (only the items in are taken into consideration when

compute the for ) which is ’s observation about ’s performance to user cluster

, . The methods for updating the average quality, stability and selection frequency in

’s peer profile , , are given below, where , , , , , are the updated

value for peer and cluster , in the three matrixes, respectively:

,, ,

, 1 

(4.2)

,   , 1 

(4.3)

,

0, , 2

, 1 ,

,

χ ,

, 1,

(4.4)

Page 132

Equation (4.2), (4.3) and (4.4) simply keep track of the average and standard-

deviation of the recommender performances as well as the number of times the

recommender peers were selected (for a user cluster). In the next section, we will

describe the proposed recommender selection approach based on these three matrixes.

4.3.4 Recommender Peer Selection

In this section, a novel technique is proposed that allows manager peers to

effectively and efficiently select contractor peers based on the proposed recommender

peer profiles described in Section 4.3.3 for assistances in making quality

recommendations. The proposed peer selection strategy is based on the famous Gittins

Indices technique (John, 1989) developed for solving the exploitation vs. exploration

problem, as such, it enables the manager peers to efficiently learn their contractor peers

as well as maintain their recommendation quality to the users.

4.3.4.1 Gittins Indices

In this section, a brief explanation of the Gittins indices is given. The Gittins

indices (John, 1989) is developed for the -armed bandit problem (which is a subset of

the exploitation vs. exploration problem) that deals with a slot machine with k arms. An

amount of reward will be given when an arm is pulled. However, in each period, only a

limited number of arms can be pulled (normally one arm). Different arms have different

reward distributions, and the reward distributions for the arms are initially unknown. The

objective is to choose which arms to pull that will maximise the total rewards over time

based on previous experience and obtained rewards. Formally, the k-armed bandit

problem is to schedule a sequence of pulls maximising the expected present values of

Page 133

(4.5)

where indicates the time points, denotes the sum of the rewards obtained by

pulling a set of arms at , and is a fixed discount factor where 0 1.

Traditionally, dynamic programming was the preferred framework for solving

the bandit problem. It requires analysis of all possible combinations of the pulling

sequences. However, Gittins developed a solution in 1972 that required computation

only on the current states of the individual arms. Gittins suggests comparing each

potential action (i.e. a pull) against a reference arm with a known and constant reward¸

instead of to compare all possible actions against each other (John, 1989). Gittins proved

it is optimal to select actions with expected rewards equal to the reference actions with

the highest equivalent rewards (i.e. Gittins index values) for each pull (John, 1989).

Specifically, a Gittins index value of an arm is computed based on the average

and standard deviation of the rewards generated from the arm as well as the number of

times the arm has been pulled. The application of the Gittins indices for solving the

multi-armed bandit problem is therefore straight forward: we simply compute the Gittins

index values for every arms (based on their current average and standard deviation of the

rewards generated and the number of times each of them are pulled), and pull the arm

with the highest index value. As the arm selection task involves only the current states

of the arms (i.e. current average and standard deviation of the rewards and number of the

times being pulled), it is, therefore, both memory and computationally efficient (when

comparing to dynamic programming based solutions).

Page 134

The theorem background and the relevant index value generation techniques of

the Gittins Indices technique are detailed in (John, 1989). This thesis mainly focuses on

the application of the Gittins indices in the context of the recommender peer selection

task. In this thesis, we employed one of the Gittins methods to generate the index values

based on the multi-population sampling in relation to the mean and standard deviation

rewards of the arms. For a given discount factor , the Gittins indices can be calculated

by back-solving the recurrence relation:

, , ,  1

, , , , 1 | , ,  

(4.6)

where is the current number of trials, is the average rewards generated from

past trials, and is the standard deviation of the rewards. is the updated average

rewards giving is the new reward generated by the distributions function | , ,

in the 1 trial, such that

1

and denotes the updated standard deviation of the 1 rewards

1 1

Generally, Equation (4.6) expresses the selection between a referenced arm

with a constant reward and an uncertain arm with an expected reward . In the

Equation (4.6), the term

, , , 1 | , ,

Page 135

indicates that the reward obtained from the next selection (i.e.   1) will be discounted

by . Similarly, the left term in the maximum function in Equation (4.6), , is the

cumulative reward for always choosing the referenced arm (with the constant reward ).

Therefore, the Gittins index of a given arm is a value of that makes both the first and

the second arguments of the maximum function in Equation (4.6) to be equal (Azoulay-

Schwartz et al., 2004, John, 1989).

Figure 4.5. The relation between  and Gittins Indices when 0.9

Given an arm which has been pulled for times, and generated an average

reward and a standard deviation , Gittins denotes the index value for the arm as

, , , and he also proved in (John, 1989) that:

, , 0,1,  

(4.7)

where 0,1, is the index value for an arm being pulled for times with a zero

average reward and a standard deviation of 1. Gittins has calculated the value of

0,1, for different combination of and in (John, 1989). Table 4.2 lists the Gittins

indices (i.e. 0,1, ) for 0.9, and this table is calculated by combining ‘table 1:

-1

0

1

2

3

4

5

6

1 10 100 1000

v (0

,1,n

)

n

Page 136

normal reward process with a known variance’ and ‘table 3: the ratio of indices for cases

of unknown variance and known variance’ in the appendix of (John, 1989). We also

illustrate the relation between and 0,1, in Figure 4.5, which is portrayed based on

the Table 4.2.

Table 4.2. The Gittins indices table for 0.9

0,1, 0,1, 0,1,

2 5.169212 20 0.074436 200 0.007931

3 0.735712 30 0.050491 300 0.005307

4 0.416059 40 0.038287 400 0.003988

5 0.30622 50 0.03086 500 0.003194

6 0.246668 60 0.025856 600 0.002664

7 0.208662 70 0.022254 700 0.002285

8 0.181654 80 0.019534 800 0.002

9 0.161279 90 0.017409 900 0.001778

10 0.144795 100 0.015701 1000 0.001601

Based on Equation (4.7), it can be observed that as an arm’s average rewards

increases, its index value increases too. Moreover, despite the average rewards, the

standard deviation of the arm’s past performances and the number of times the arm has

been pulled also play important roles in the index calculation. It can be seen from Figure

4.5 that the standard index value 0,1, is only significant when is small (i.e.

  3 ), when gets bigger, 0,1, shrinks. By combining 0,1, with as

shown in Equation (4.7), the contribution of the standard deviation of an arm’s past

rewards to the index value , , decreases drastically when increases. Intuitively,

Equation (4.7) indicates that when our experience to an arm is low (i.e. is small) it is

better to select the arm if it is highly risky (i.e. if is higher), because the risky arm

Page 137

might potentially generate high rewards in the future. In contrast, if we already have a

long experience with the arm, then it would be more important to look at the arm’s

average rewards rather than to gamble on its instability.

The above concepts can be adapted into the recommender peer selection problem.

Recommender peers can be treated as the arms in the armed bandit problem. The

number of times a recommender has been chosen corresponds to the number of times an

arm has been pulled. The calculations for and are the same to the profile

updating to the average and the standard deviation of the recommender peer

performance (i.e. Equation (4.2) and Equation (4.4)). Initially, if a recommender peer (i.e.

manager peer) does not know about other peers (i.e. contractor peer) very well (i.e. low

values in the selection frequency matrix), then it would be a good strategy to select peers

with lower stability, because the unstable peers might become better in the future

(whereas stable peers stay unchanged). However, after a certain period, the stability of

the peers becomes insignificant, because as the number of trials increases the average

performance of the peers become reliable and dominate over the stability.

4.3.4.2 Selection Strategy for EDRS

Based on Section 4.3.4.1, when a manager peer wants to find a best

contractor peer to make a recommendation to a user , where , Φ and

, , the following equation is used to select the most suitable peer:

argmax\

, , ,  

(4.8)

Page 138

where , is the Gittins index function that maps , (i.e. selection frequency) to the

corresponding 0,1, , based on Table 4.2. In Equation (4.8), firstly calculates the

average performance, stability and selection frequency of the available peers to the user

cluster that belongs to (i.e. , ). Then computes the index values for every peer

based on Equation (4.7). Finally, the most preferred peer will be the one that has the

highest index value. By setting up a cut-off for the index value, multiple recommender

peers with index values higher than the cut-off can be selected. However, selecting

multiple peers to make a recommendation requires recommendation fusion that will be

briefly discussed in latter sections. In addition, the discount factor as depicted in

Equation (4.5) and (4.6) discounts the future rewords exponentially, this implies that it is

more important for a recommender to achieve higher performance in the present rather

than to achieve the same performance in the future. Therefore, the smaller the value of ,

the severer the future rewards are discounted. In this thesis, we suggest a large value for

(i.e. 0.9) which discounts the future rewards in a gentle fashion, because we

perceive that the long term relationships between the recommenders are necessary.

4.3.4.3 An Example

In this section, an example is provided to demonstrate the proposed

recommender peer selection method. We start by assuming that a recommender

(manager peer) has made recommendations to a user in cluster , by consulting

four contractor peers , , and before, the past performances of the

recommender peers are computed based on Equation (4.1) and are given in Table 4.3,

where the number of times that , , and are selected to make recommendation.

Page 139

Table 4.3. Performance histories for four recommender peers

Peers Rewards (χ) Received

0.2, 0.6, 0.3

0.3, 0.45, 0.42

0.9, 0.4, 0.8, 0.6

0.7, 0.7, 0.8, 0.75, 0.68, 0.8

Given , as the user cluster, ’s profiles to the peers are three 4-dimension

column vectors:

0.3667, 0.39, 0.675, 0.7383

0.2082, 0.0794, 0.2217, 0.0531

3, 3, 4, 6

Apparently, ’s profile vector is the elements of the vectors, that is , , , ,

, . The vectors are calculated by applying Equation (4.2), (4.3) and (4.4) to Table 4.3.

In order to calculate the Gittins indices for the recommender peers, we have to convert

into standard Gittins indices as described in Section 4.3.4.2:

0.735712, 0.735712, 0.416059, 0.246668

The conversion from to is simply a table lookup to Table 4.2. Next, we

compute the intended Gittins indices vector, , by combining , and based

on Equation (4.7):

0.5918, 0.4484, 0.7673, 0.7514

where denotes element-wise multiplication.

If only the past performances of the recommender peers are considered, is the

best choice because it performed best (i.e. , max 0.7383) in the past.

Page 140

However, based on , it suggests that is the best choice. We can better understand

the rationale behind the choice by comparing the stability between and . Even

though averagely performed better than (i.e. , , ), it is still worthwhile to

take risk on , because has only been selected for 4 times and its performances

varied drastically ( , 0.2217). might be a still a good choice (with the second

highest index , 0.7514), however, because it is already relatively stable ( ,

0.0531), we can take chance to learn more about other peers first. Therefore, is

preferred to . The same concept can be applied when comparing with . With the

same selection frequency ( , , 3), although generally outperforms , it is

suggested to take risk on the unstable peer , as it might potentially improves its

performance.

4.4 RECOMMENDATION MERGE

Recommendation merge is an important task for distributed recommender

systems, but it is not a key focus of this thesis due to time limitations. In this section, we

present a simple technique for a manager peer to merge recommendations generated

from a set of contractor peers. Here, we assume the set of contractor peers is selected

based on the peer profiling and selection strategy described in Section 4.3, and therefore,

each of these selected peer will be associated with a Gittins score (see Section 4.3.4.2)

that indicates the expected utility the manager peer might obtain when the contractor

peer’s recommendation is adopted.

In the recommendation merge task, a manager peer selects top contractor

peers Φ , , . . with the highest Gittins scores for recommendation

suggestions, and each of these selected contractor peers Φ sends the manager peer

Page 141

a recommendation list , , , , , , , , … , , , , . As discussed in

Section 4.3.3, only items that are in are considered by . Therefore, after removing

the items from which are not in , for each value pair , , , in , , is

the recommended item and , 0,1 indicates ’s confidence that , will be

preferred by the manager peer’s target user. Note, because different recommender peers

might have different recommendation methods, so their confidence scores (i.e. , )

might not be directly comparable to each other. For simplicity, we assume that the

confidence scores from different recommender peers are normalised and comparable, so

that if the scores , and , of two peers , Φ to items , and , are

similar (i.e. , , ), the two peers have similar confidences to the two items.

Let , , … , be the set of items each of which is recommended by at

least one of the contractor peers, i.e. the items in , must appear at least in one of the

recommended lists , , , , , , … , , , , 1, . . . If 'ppi was

not recommended by contractor , then , 0 . Merging the recommended lists

returned from contractors , , . . can be viewed as to recalculate the scores to the

items in , based on the scores given by the contractors. Let

, , , , … , , be the final recommendation list after merging the

recommended lists  , , , , , , … , , , , 1, . . , to merge the

recommendations is to calculate the scores based on , ,  1, . . , 1, . . .

As the contractor peers Φ are selected by manager peer based on the

Gittins scores, therefore, each is associated with a corresponding Gittins score ,

(see Section 4.3.4.2 and Section 4.3.4.3 for detailed Gittins score computation). Note,

the symbol , is borrowed directly from Section 4.3.4.3, where indicates , , the

target user’s belonging cluster in . We propose to use a linear combination of the

Page 142

contractor peers’ Gittins scores ( , ) and the recommendation scores (i.e. , ) they

assigned to the items to calculate the final score to the items. The algorithm to perform

the merging is given below:

Algorithm 4.1 _ ,

Input , , , , … , is the set of Gittins scores for the selected

contractor peers. , denotes the Gittins score the manager peer assigned

to the contractor peer Φ for its recommendation to the target user in

, .

, , … is the set of recommendation lists generated from the

selected recommender peers. denotes the recommendation generated by

the contractor peer Φ to the manager peer .

Output is the merged recommendation list

1) SET , , an initially empty set for storing all items involved in

2) FOR EACH

3) FOR EACH , , ,

4) SET , ,,

5) END FOR

6) END FOR

7) SET , an initially empty set for storing the final merged recommendation

list

8) FOR EACH ,

9) SET 0, 0

10) FOR EACH

Page 143

11) SET , ,

12) IF , 0 THEN ,

13) END FOR

14) SET /

15) SET ,

16) END FOR

17) Return as the merged recommendation list

In Algorithm 4.1, the manager peer firstly finds all items that are

recommended by the contractor peers (i.e. from line 1 to line 6) and stores them in the

candidate item set ,. As it is possible that different contractor peers may suggest the

same items in their recommendation lists, the size of , the candidate item set is,

therefore, between the size of the largest recommendation list (in the case that all

contractor peers recommend the same set of items) and the size of the union of all

recommendation lists (in the case that all contractor peers recommend different items),

specifically:

max | | | ,| | |

In Algorithm 4.1, the linear combination of the Gittins scores and the

recommendation item scores is implemented from line 10 to 13. Line 11 indicates that

the item scores received from the contractor peers are factored by the peers’ Gittins

scores. Thus, items suggested by the contractor peers with higher Gittins scores will

receive high final scores. Moreover, in line 14 of Algorithm 4.1, is normalised by the

sum of the Gittins scores of the contractor peers who have recommended the item.

Page 144

4.5 EXPERIMENTS AND EVALUATION

As mentioned in the beginning of this chapter, the major objective of the research

presented in this chapter is to demonstrate the possibility of alleviating the cold-start

problem by enriching the information resources with help from recommenders of other

parties. Specifically, we proposed an EDRS framework (see Section 4.2) for modelling

the interactions and communications of the recommenders, and the goal of the

framework is to allow the recommenders to improve their recommendation quality by

integrating their recommendations together. In order to facilitate the interaction protocol

of the proposed EDRS, in Section 4.3 we proposed a recommender peer profiling and

selection technique which allows recommenders to effectively learn from each other and

select the partner recommenders that can help them best. Based on the goals we

presented in this chapter, the experiments we conducted in this part of thesis aim to

verify the following:

Whether recommenders can improve their recommendation quality as well

as their resistance to the cold-start problem by incorporating aid from

recommenders of other organisations.

Whether the proposed peer profiling and selection strategy can effectively

facilitate the interactions of the recommender peers within the proposed

EDRS framework.

In this experimentation, multiple recommenders with different capability in

making recommendations are constructed, and we allow them to interact with each other

based on the proposed EDRS framework. Essentially, these recommenders employ the

proposed peer profiling and selection strategy presented in Section 4.3 to learn from and

select each other in order to improve their recommendation making. Our main focus is to

examine whether incorporating aid from other recommenders can indeed improve

Page 145

recommenders’ recommendation quality and also to evaluate the effectiveness of the

proposed profiling and selection strategy.

Note, due to the limited scope of this thesis and also the recommendation merge

is not a key focus of this thesis, the experiment is configured such that the manager peers

select only one contractor peer for each recommendation making round, and the

manager peer forward directly the recommendations from the selected contractor peer to

the target user. Hence, the recommendation merge technique described in Section 4.4 is

not involved in this experiment. However, as the peer profiling and selection strategy is

the most essential part of the proposed EDRS framework, our experiments sufficiently

cover the two previously mentioned experimentation goals.

In Section 4.5.1, the dataset we employed for the experiments is discussed. In

Section 4.5.2, the experiment process and settings used for evaluating the proposed peer

profiling and selection technique are discussed. Finally, in Section 4.5.3 the experimental

results are presented and explained.

4.5.1 Data Acquisition

The dataset employed in this experiment is the ‘Book-Crossing’ dataset

(http://www.informatik.uni-freiburg.de/~cziegler/BX/) which is also the main

experiment dataset employed in Chapter 3. Please refer to Section 3.3.1 for more details

about the dataset.

As this experiment involves only the standard item-based collaborative filtering

recommender, the product taxonomy data employed in Chapter 3 is not used in this

experiment.

Page 146

4.5.2 Experiment Setup

As the main purpose of this experiment is to evaluate the proposed interaction

protocol and the peer profiling and selection technique (rather than evaluating a new

recommendation technique or algorithm) in a distributed recommender system, therefore,

the overall setup of this experiment is different from the setup for non-distributed

recommender systems.

In this experiment, it is required to simulate the interactions (i.e. profiling and

selection) among the recommenders from different organisations, and therefore the first

step in the experiment setup process is to construct multiple recommenders with

different capabilities and underlying knowledgebase (i.e. datasets). Next, the testing

dataset is constructed for evaluating the recommenders’ recommendation quality.

Importantly, the recommendation quality comparison between recommenders utilising

the proposed EDRS framework (i.e. getting aid from other recommenders) and stand-

alone recommenders (i.e. making recommendations based on their own efforts) are

carried out. Moreover, the effectiveness of the proposed peer profiling and selection

technique is also examined by comparing it with other peer selection strategies. Note, the

proposed peer profiling strategy requires the manager peers to get user feedbacks for all

of their recommendations (see Section 4.3.3) so they can determine their contractor peers

performances based on the feedbacks and then update their peer profiles. Hence, it is

necessary to provide a way to allow the user feedbacks in the experiment. The tasks

involved in this experiment setup are detailed in the following subsections.

4.5.2.1 Constructing the Recommender Peers

In this experiment, four recommenders of different organisations are constructed

to simulate the proposed recommender peer interactions. These four recommenders are

Page 147

named as ORG1, ORG2, ORG3 and ORG4, and they are equipped with different

datasets but use the same underlying recommendation technique.

By evaluating the performances of the recommenders with the same

recommendation technique and different underlying datasets, we can evaluate the

performance of the recommenders based on their available information resources (i.e.

their underlying datasets and collaboration from other recommender peers) without the

impact from using different recommendation techniques. Moreover, the results from the

experiments can also be used to verify the proposed solution to the cold-start problem

(i.e. enriching the information resources from other parties).

The recommendation technique employed by the four recommenders is the

standard item-based collaborative filtering technique that is identical to the benchmark

recommender IR employed in Chapter 3 (see Section 3.3.3.1 and Section 2.1.2.1 for

more details). The use of the state-of-the-art recommendation technique ensures that our

experiment can be compared and verified with other studies. Moreover, it also suggested

that the proposed EDRS framework and peer profiling and selection strategy can be

easily adopted by existing recommenders.

The main differences among the four recommenders are in their underlying

datasets, specifically, they all have different customer sets (or user sets). We firstly select

6500 users from the Book-Crossing Dataset and then cluster them into 20 user clusters

based on their item preferences (i.e. explicit item ratings). We denote the overall user set

as and the 20 user clusters as , , … , . Specifically, | | 6500 ,

and .

From these 6500 users in , 5000 users are selected as the training user set

(i.e. for forming the underlying datasets of the recommender peers) and the rest of 1500

users then forms the testing user set , where and . Furthermore,

Page 148

we denote the set of training users within cluster as (i.e. ), the set

of testing users within cluster as (i.e. ), and .

Importantly, the users in are divided into the clusters first, and the 1500 users in the

testing set are then selected from each of the clusters. This process allows us to keep

track of the percentages of the different user types (i.e. users in different clusters) in the

testing user set. The allocation details for are shown in Table 4.4. Specifically, each

row in Table 4.4 shows the user allocation detail for a cluster. For example, the first row

in the table shows that there are totally 2278 users being grouped into among which

300 users are selected into the testing user set (i.e. | | 300). Note, because the size

of clusters ,  ,  , and , we do not select any users from them into the

testing set.

Next, the datasets for the four recommender peers ORG1, ORG2, ORG3 and

ORG4 are constructed from , and they are denoted as ,  ,  , and respectively.

As mentioned in Section 4.2, the users and recommender peers in the proposed EDRS

are in many-to-many relation, thus, it is possible that a user can exist in multiple

recommenders’ datasets (i.e. ). Table 4.5 shows the detailed

allocations for the datasets of the four recommenders. The four recommenders are each

having different user set sizes, ORG1 has the largest dataset with 2000 users, ORG4 has

the smallest dataset with 700 users, and ORG2 and ORG3 both have 1250 users in their

datasets. Even the total number of users involved in the four user sets is 5200 according

to Table 4.5, the total number of users involved in these four recommenders is actually

equal or smaller than 5000 (i.e.  ) due to the user overlapping

allowed among the datasets. Moreover, it is shown in that different recommenders have

different numbers of users in different clusters, for example, ORG1 has the highest

number of users in (i.e. | | 1500) whereas ORG3 has the highest number

Page 149

of users in (i.e. | | 150). This setting allows us to examine whether

different user sets will affect the recommenders’ performances for different users.

Table 4.4. Allocation details for the training and testing user sets

Cluster Training Set ( ) Testing Set ( ) Total

2078 (91.2%) 200 (8.8%) 2278

173 (63.3%) 100 (36.7%) 273

11 (100%) 0 (0%) 11

4 (100%) 0 (0%) 4

230 (69.7%) 100 (30.3%) 330

188 (55.6%) 150 (44.4%) 338

82 (62.1%) 50 (37.9%) 132

230 (69.7%) 100 (30.3%) 330

156 (75.7%) 50 (24.3%) 206

229 (69.6%) 100 (30.4%) 329

237 (61.2%) 150 (38.8%) 387

216 (68.4%) 100 (31.6%) 316

123 (100%) 0 (0%) 123

77 (100%) 0 (0%) 77

18 (100%) 0 (0%) 18

174 (53.7%) 150 (46.3%) 324

214 (58.8%) 150 (41.2%) 364

154 (75.5%) 50 (24.5%) 204

247 (71.2%) 100 (28.8%) 347

59 (54.1%) 50 (45.9%) 109

Total 5000 (76.9%) 1500 (23.1%) 6500

Page 150

Table 4.5. Dataset allocation details for the four recommender peers

Cluster ORG 1 ( ) ORG 2 ( ) ORG 3 ( ) ORG 4 ( )

1500 500 250 0

100 100 0 0

0 0 0 0

0 0 0 0

10 150 0 200

90 100 0 0

0 0 0 50

30 0 0 0

0 50 150 0

0 0 0 200

0 150 0 200

100 0 200 0

0 0 0 0

0 0 0 0

0 0 0 0

60 0 250 0

60 0 250 0

0 0 150 0

50 200 0 0

0 0 0 50

Total 2000 1250 1250 700

Page 151

4.5.2.2 Evaluation Metrics

The classification accuracy metrics (i.e. Precision, Recall and F1 metrics) are

chosen for the performance evaluation of the recommenders against the users in the

testing user set. As these metrics have also been used for the experiments in Chapter 3,

please refer to Section 3.3.3.3 for detailed explanations to these metrics.

As described in Section 2.4 and Section 3.3.3.3, the classification accuracy

metrics are mainly based on comparing the recommended item list and the set of user

preferred items. In this experiment, for each testing user , we divide the set of

items explicitly rated by (denoted as , where T) into two halves denoted by

and , where , and | | | |. As all of the items in are

explicitly rated by , therefore, for any item , is associated with a numeric

item rating , 0,1 . For the two item sets , , and the associated item

ratings are used to represent ’s user profile (i.e. the recommenders make

recommendations to based on ’s ratings to the items in ), and the items in ,

conversely, are used to form the user preferred item list for evaluating the

recommendations made to . However, not all the items in are preferred by the user

. The items with low rating values should not be considered as the user’s preferred

items because has specifically indicated that they are disliked. Hence, the final testing

item set is constructed by removing all items with ratings below ’s average rating

from .

For evaluating the recommenders’ recommendation quality to a given testing

user , the recommenders are firstly provided with ’s profile (i.e. and the

associated ratings), then the recommenders generate their recommendations to , finally,

Page 152

the recommendations generated from the recommenders are evaluated against the testing

item set by utilising the classification accuracy metrics (i.e. Precision, Recall and F1).

4.5.2.3 Benchmarks for the Peer Profiling and Selection Strategy

As mentioned earlier, one of the objectives of this experiment is to evaluate the

effectiveness of the proposed peer profiling and selection technique described in Section

4.3. Hence, it is important to include other profiling and selection techniques as baselines

in order to conclude the significance of the proposed technique. However, to the best of

our knowledge, there are no other existing studies available for the recommender peer

profiling and selection tasks required for the proposed EDRS (the concept of EDRS is

new and firstly proposed by this thesis). As there are no existing standard baseline

techniques available in distributed recommender systems, we therefore have adapted

techniques from other research domains that are reasonably applicable to the required

peer profiling and selection task. In this experiment, the following three peer profiling

and selection strategies are compared:

Gittins: The proposed recommender peer profiling and selection technique

as described in Section 4.3.

BPP: Best Past Performances. It is the most fundamental and intuitive

strategy being used for the profiling and selection related tasks in many

research domains (e.g. the collection selection task in distributed

information retrieval (Kretser et al., 1998) ) . The basic idea behinds BPP is

to select recommender peers with the best average past performances to the

target users’ belonging clusters. Specifically, a BPP based recommender

peer Φ profiles other peers with only the peer average performance

Page 153

matrix (see Section 4.3.3), and it finds the best contractor peer for

making recommendations to a target user (where , ) by:

argmax\

,

BPP is different from Gittins as it does not take the peer stability (i.e. ) and

selection frequency into considerations.

Rand: The manager peers based on this strategy keep no knowledge about

other peers and select contractor peers at random. This strategy is included

in this experiment to show the significance of having a reasonable peer

profiling and selection strategy in the proposed EDRS.

Gittins_NC: This selection strategy is a simplified version of the proposed

strategy Gittins. Essentially, Gittins_NC assumes all users belong to one

cluster. Even Gittins_NC still profiles recommender peers based on their

average performance, stability and selection frequency, and the selection is

also based on the combined Gittins scores as described in Section 4.3.4.2, it

does not profile the recommender peers by considering the performance

differences for users in different clusters.

BPP_NC: Similar to Gittins_NC, this profiling and selection strategy does

not differentiate peers’ performance differences for users in different clusters,

and it employs only the average past performances of the recommender

peers to select (i.e. as similar to BPP). The main purpose of having

Gittins_NC and BPP_NC included in this experiment is to demonstrate

empirically that different recommenders have different performances

towards users in different clusters.

Page 154

4.5.2.4 Simulating the User Feedbacks

It is described in Section 4.3.3, the manager peers learn and profile the contractor

peers based on the target users’ feedbacks to their recommendations. As there are no real

users involved in this experiment, therefore, we need to simulate the user feedbacks to

the recommenders in order to evaluate the proposed peer profile and selection technique.

As stated in Section 4.5.2.2, a testing user ’s rating data is divided into

two parts, is for training and is for testing. Hence, for a set of items T

recommended by a recommender peer to , we can use ’s real ratings to the items in

as the feedbacks to as ’s explicit ratings to items in are directly available.

However, for those recommended items that are not in (i.e. \ ), the true user

feedbacks from are not available. In order to supply feedbacks for those items in

\ , we have constructed a feedback simulator that makes feedbacks by predicting

users’ true ratings. The feedback simulator utilised the standard collaborative filtering

technique (described in Section 2.1.2) to predict a target user’s ratings based on the

entire user dataset as its knowledgebase, the complete target user profile (i.e. ).

As the term ‘simulation’ suggested, the simulated feedbacks for to items in

\ are not as accurate as ’s true ratings (i.e. ratings for these items in ).

However, the simulated feedbacks can be considered more close to the user’s true ratings

than the recommendations made by all the recommender peers in the experiment,

because of the following reasons:

The entire user set is employed by the feedback simulator as the base to

make item rating prediction, whereas the recommendations generated from

the four recommenders in this experiment are only based on small subsets of

. For example, the simulator simulates a user ’s feedbacks based

on other 338 similar-mind users (see Table 4.4), whereas the recommender

Page 155

in ORG2 has only 100 similar-mind users to be based on for making

recommendations to .

The simulated feedbacks for a testing user is based on his or her

complete past rating data , whereas all the recommender peers in the

experiment make the recommendations to based on only half of the

complete rating data (i.e. /2).

Even though the feedbacks generated by the simulator may not exactly the same

as the user’s true ratings, the combination of the user’s true ratings to the items in

and the simulated feedbacks to the items in \ ensures that the manager peers are able

to judge the contractor peers’ performance at a reasonable level, and it is sufficient for

the purpose of this experiment.

4.5.3 Experimental Results

In this section, the results obtained from the experiment are presented and

discussed.

Each of the four stand-alone recommenders (i.e. ORG1, ORG2, ORG4 and

ORG4) can run by itself using its own dataset. However, the performance of the

individual recommenders may not be satisfactory due to the insufficiency of the dataset.

The EDRS framework proposed in this thesis can improve the performances of all

involved participant recommenders by allowing them to share datasets and

recommendations. Therefore, it is expected that the distributed recommendation system

with a reasonable peer selection strategy outperform the individual recommenders.

Figure 4.6, Figure 4.7 and Figure 4.8 present the precision, recall and F1 results obtained

from running the four stand-alone recommenders (i.e. ORG1, ORG2, ORG4 and

Page 156

ORG4) and the distributed recommendation system with five peer selection strategies

described in Section 4.5.2.3 (i.e. Rand, BPP_NC, Gittins_NC, BPP and Gittins),

respectively.

Let us firstly look at the performance of the distributed recommender system

with the five different profiling and selection strategies (i.e. Rand, BPP_NC, Gittins_NC,

BPP and Gittins). Among these five strategies, Rand is the only strategy that does not

have profiles for the recommender peers, and it randomly selects peers for making

recommendations. Based on the experiment results shown in Figure 4.6, Figure 4.7 and

Figure 4.8, Rand performed the worst among all of the five strategies, and it even

performed worse than two of the stand-alone recommenders ORG3 and ORG4, which

make recommendations only based on their own datasets. In contrast, the other four

strategies (i.e. BPP_NC, Gittins_NC, BPP and Gittins) that profile recommender peers

based on the peers’ past performances and select peers’ based on their profiles all

achieved much better results than all stand-alone recommenders except for ORG3.

Since ORG3 is the best performed stand-alone recommender and therefore very often

selected by the manager recommender, the distributed system with some of these

strategies achieved similar performance as what ORG3 does. This result suggests that

by sharing datasets and selecting the most appropriate recommender to make

recommendations, the distributed recommendation system can greatly improve

recommendation quality. Particularly, for those peers which suffer from the cold-start

problem (such as ORG1 and ORG2), the amount of improvement is significant, for

instance, the performance of both ORG1 and ORG2 can be improved by more than 50%

if they adapt any of the four strategies to profile and select peers.

Among the four rational strategies (i.e. BPP_NC, Gittins_NC, BPP and Gittins),

BPP and Gittins profile and select peers based on their performance to users in different

Page 157

clusters. In contrast, BPP_NC and Gittins_NC do not consider the fact that different

peers might perform differently for users in different clusters and profile peers based on

their average performance over all users. As shown in Figure 4.6, Figure 4.7 and Figure

4.8, the cluster-based strategies BPP and Gittins significantly outperformed the non-

cluster-based strategies BPP_NC and Gittins_NC. This is because the cluster-based

strategies can find the best recommender peers for making recommendations based on

the target users’ belonging clusters. In contrast, BPP_NC and Gittins_NC select

recommender peers based on their average past performances to all users. Therefore,

they will select peers performed averagely best in the past despite that these peers might

be unable to produce good recommendations for some target users in certain clusters.

Finally, the experiment results show that the Gittins indices based strategies (i.e.

Gittins and Gittins_NC) performed better than that of the standard performance based

strategies (i.e. BPP and BPP_NC). Specifically, Gittins outperformed BPP and

Gittins_NC outperformed BPP_NC. This result suggests that by combining the selection

frequency and recommendation stability into peer profiling and selection process (as

discussed in Section 4.3), the best performed peers can be more accurately identified

than only based on the peers’ average past performances.

Page 158

Figure 4.6. Precision results for different recommendation settings

Figure 4.7. Recall results for different recommendation settings

0.0677 0.0622 

0.1661 

0.0862 0.0719 

0.1656 0.1674 0.1776 

0.1965 

0

0.05

0.1

0.15

0.2

0.25

Pre

cisi

on

0.2064 0.2008 

0.3461 

0.2173 0.2082 

0.3454 0.3494 0.3771 

0.4172 

0

0.05

0.1

0.15

0.2

0.25

0.3

0.35

0.4

0.45

Rec

all

Page 159

Figure 4.8. F1 results for different recommendation settings

4.6 CHAPTER SUMMARY

In this chapter, we suggested a new distributed system paradigm for

recommenders, namely, Ecommerce-oriented Distributed Recommender System

(EDRS). EDRS is designed to allow the recommenders from different organisations or

parties to share recommendations with each other, so all of them can achieve better

recommendation quality and services to their users. In addition, as the recommenders

within the proposed EDRS no longer make recommendations solely on their own efforts,

they are therefore more resistant to the cold-start problems.

In order to facilitate the interaction among the recommenders in the EDRS, a

novel peer profiling and selection strategy is proposed in this chapter. The proposed

strategy profiles and selects recommender peers based on their past recommendation

performance, stability and selection frequency in cluster level, and our experiment

results show that the proposed strategy allows recommender peers to effectively learn

from each other and select the most appropriate peers to provide satisfactory

recommendations to their users.

0.0960 0.0908 

0.2089 

0.1180 0.1015 

0.2084 0.2107 0.2258 

0.2499 

0

0.05

0.1

0.15

0.2

0.25

0.3

F1

Page 160

Chapter 5 

5Conclusions

In the last decade, many techniques have been proposed for improving

recommenders’ recommendation quality and their resistance to the cold-start problem.

Most of these existing techniques focus on exploring new ways to better utilise the

available data and information resources in order to generate better recommendations.

However, given very limited data and information resources, the amount of

improvements that can be achieved by these techniques are also limited. In this thesis, a

novel perspective is proposed for improving the recommendation quality and alleviating

the cold-start problem: enriching the available information resources for the

recommenders. Two novel strategies are presented in this thesis to achieve the

information resource enrichment. The first strategy is to consider other facets of the data

and information resources. Specifically, a novel taxonomy-based recommender system,

HTR, is developed in this research. It is able to mine personal-related user taxonomic

preference information from non-personal related product taxonomic descriptors, and it

then combines this new information resource (user taxonomic preferences) with the

available user rating data to generate recommendations (see Chapter 3). The second

strategy for the information resource enrichment is by gathering information resources

from other parties. An Ecommerce-oriented Distributed Recommender System (EDRS)

is presented in this thesis that allows information resources and recommendations to be

shared by multiple recommenders, and they are then able to utilise the shared

recommendations and resources to generate better recommendations (see Chapter 4).

The techniques presented in this thesis are evaluated with popular experimental datasets

Page 161

(Book Crossing Dataset (http://www.informatik.uni-freiburg.de/~cziegler/BX/)) and

standard recommender framework (Tastes (http://taste.sourceforge.net/)) to ensure the

soundness of the experimental results. The results show that the proposed HTR and

EDRS are able to produce high quality recommendations even in the case of the cold-

start situations.

Section 5.1 presents the main contributions of this research. Section 5.2 discusses

the possible directions for the future work in the area of this research.

5.1 CONTRIBUTIONS

The contributions made by this research are listed below:

Discovering the item preference to item taxonomic preference relation:

In this thesis, the implicit relationship between users’ item preferences and

item taxonomic preferences is investigated. This relationship states that

users share similar item preferences might also share similar item taxonomic

preferences. A novel technique is proposed to efficiently and effectively

mine and extract this relation/knowledge from the combination of user

rating data and product taxonomic descriptors. Additionally, the soundness

of the relationship between user item preference and item taxonomic

preference is also empirically evaluated. The details can be found in Section

3.2.3 and Section 3.3.2.

A novel taxonomy-based recommender system: Based on the proposed

relationship between item preference and item taxonomic preference, a

novel recommender system, HTR, is proposed. HTR is very competitive in

terms of computation efficiency and recommendation quality, and most

importantly, it is able to produce high quality recommendations under

Page 162

severe cold-start situations. The details of HTR and related studies can be

found in Section 3.2 and Section 3.3.3.

A novel distributed recommender system: This thesis suggested that one

of the possible ways to achieve the information enrichment is to obtain

resources from other parties. This research proposed a novel distributed

recommender system namely, EDRS, and it allows recommenders from

different parties to share their recommendation and information resources

with each other to enhance their recommendation quality. The background

rationale, interaction protocol, system infrastructure and design aspects of

the proposed EDRS are comprehensively reviewed and presented in this

thesis. The details can be found in Section 4.2.

A novel recommender peer profiling and selection strategy: In order to

enhance the overall performance of the EDRS, a novel peer profiling and

selection strategy is proposed in this thesis. The proposed strategy profiles

and selects recommender peers based on their average performance,

performance stability and selection frequency, and it allows recommenders

to efficiently learn about each other and choose the most effective peers to

assist in making recommendations. It is shown in the experiments presented

in this thesis, by adopting the proposed profiling and selection strategy the

performance of the EDRS is effectively improved. The related information

and experiments can be found in Section 4.3 and Section 4.5.

Three novel neighbourhood formation related techniques: In addition to

the main contributions of this thesis (i.e. HTR and EDRS), three new

recommender-related techniques are also developed during the research, and

they are:

Page 163

o A novel similarity measure – Statistical Attribute Distance (SAD). It

allows user profile similarity to be more objectively measured by

considering the popularity differences among the attribute values in the

user profiles. The detailed information for SAD is described in Appendix

A.

o A novel clustering algorithm – Hybrid Partitional Clustering (HPC). It

features in its efficiency, accuracy and the ability to produce

automatically optimal cluster partition without involving complicated

manual configurations. The detailed information for HPC is described in

Appendix B.

o A novel neighbourhood estimation technique – Relative Distance

Filtering, (RDF). RDF features in its competitive computation efficiency

and low memory requirement. The detailed information for RDF is

described in Appendix C.

5.2 FUTURE WORK

The concept of information enrichment for recommender systems proposed in

this thesis is general, and there can be many other possible ways to achieve it besides the

two strategies (i.e. HTR and EDRS) presented here. Therefore, it is one of the future

studies to investigate other new strategies to achieve the information enrichment for

recommender systems.

The HTR system presented in this thesis is specifically designed for tree structure

based item taxonomy. Indeed, such taxonomy structure has been widely used by many

ecommerce sites and applications for representing and describing item contents.

However, there are still many other item representation techniques available, and some

Page 164

of them are achieving vast popularity nowadays (e.g. item tags). Hence, it would be a

promising future work to improve the proposed HTR so it can accommodate other item

taxonomy structures or presentations.

The focus of the EDRS in this thesis is on constructing the overall framework

concept, interaction protocol and peer learning strategies. Hence, many detailed aspects

and related techniques are not covered in this thesis. In the future, the proposed EDRS

can be further improved by considering the following works:

While a novel peer profiling and selection technique is presented in this

thesis to allow manager peers to learn about contractor peers, it is desirable

to have learning strategies for contractor peers to learn about the manager

peers. By allowing manager peers and contractor peers to learn from each

other (currently only manager peers are able to profile contractor peers), the

cooperation among the recommender peers can become more effective, and

the performance of the recommender peers can be further improved.

The recommendation merging technique presented in this thesis is rather

trivial, and it can be improved or replaced by more advanced techniques.

Page 165

Appendix A: Statistical Attribute Distance

As described in Section 2.1.2, the basic idea behinds the collaborative filtering

technique is to predict the target user’s item preferences based on the tastes of other

similar minded users. Hence, it can be easily observed that determining similar minded

users for a given target user is one of the most essential parts of the collaborative filtering

based recommenders. Generally, cosine similarity and Euclidean distance are considered

as the two most popularly used similarity measures to determine the degree of similarity

between two user profiles. Assuming that the user profiles are the users’ item

preferences (i.e. item ratings), the following equations are used to calculate the similarity

between two users:

For the cosine similarity measure:

,∑ , ,

∑ , ∑ ,

(A.1)

For the Euclidean distance measure:

, , ,  

(A.2)

In both equation (A.1) and (A.2), , are the two users, , 0,1

denotes ’s explicit rating value to item . Moreover, is the set of items

Page 166

that have been explicitly rated by . For more details about the notation, please refer to

Section 3.2.1. Even though the Euclidean distance and cosine similarity measure are

simple and intuitive, they can still be improved in many aspects to better measure user

similarities. For example, the Pearson correlation coefficient measure (see Equation (3.1)

and Section 3.2.2) is often considered as a better alternative than both of these two

methods as it is able to accommodate the differences among users’ rating habits (Breese

et al., 1998, Herlocker et al., 2002, Montaner et al., 2003).

The Inverse User Frequency (IUF) method proposed by Breese (Breese et al.,

1998) was reported to even out perform the Pearson correlation coefficient measure. The

basic rational behinds IUF is to reduce the weights on universally preferred items when

calculating the user preferences to items, because these items are generally considered

less capable of capturing user similarities than uncommon items (Breese et al., 1998).

The proposed Statistic Attribute Distance (SAD) takes the concept of IUF further

by distinguishing the unpopular rating values from the popular rating values. Specifically,

while IUF considers the popularity of each item, the proposed SAD method suggests that

the popularity of each individual rating values rather than each individual item is a better

factor to be considered in similarity computation. The major limitation of IUF is that it is

strongly dependent to the completeness of the dataset and the way the dataset is

constructed (i.e. if the dataset is constructed based on sampling, we need to ensure that

the popularity distribution of the items in the sampled dataset is similar to the original

dataset). In contrast, the popularity of the item ratings in the dataset is less sensitive to

the completeness of the dataset and it allows the proposed SAD to perform a more stable

manner than the IUF.

A.1. MEMORY-BASED COLLABORATIVE FILTERING

Page 167

Memory-based collaborative filtering is the most common type of collaborative

filtering, and it is very intuitive and simple to implement (Breese et al., 1998). In this

section, the existing and proposed similarity measures will be discussed and investigated

in the context of the memory-based collaborative filtering. A typical form of the

memory-based collaborative filtering technique is listed below:

, ,\

,  

(A.3)

In Equation (A.3), denotes all users in the dataset who have previously

rated item , and , represents the predicted rating of the target user to

item . and are the average item ratings of the users and respectively. ,

denotes the actual past rating gave to . , is the user similarity measure for

computing the preference similarity between and . Finally, is a normalising

factor such that the values of the weights sum to unity. Based on the equation depicted

above, it can be observed that the accuracy of the predication is strongly dependant on

the computation of the user similarity , between the target user and all other

users  who has previously rated . There are many existing techniques can be

employed as the user similarity measure , , some of these techniques are described

previously such as cosine similarity (Equation (A.1)), Euclidean distance (Equation

(A.2)) and Pearson correlation coefficient (Equation (3.1)). Despite the many possible

implementations of , , they all have identical underlying concept: generating a high

value when and have very similar preferences, and a low value if they have no

Page 168

common tastes. In such way, the predicted rating , will approach the actual ratings to

given by the similar minded neighbours of .

A.2. INVERSE USER FREQUENCY

As mentioned earlier, there are still some advanced similarity measures other

than the standard ones such as cosine similarity, Pearson correlation coefficient, etc., and

many of these advanced similarity measures are reported to have better performances

than the standard ones (Breese et al., 1998, Herlocker et al., 2002, Montaner et al., 2003).

Inverse User Frequency (IUF) proposed by Breese et al. (1998) is one of the most known

advanced similarity measures. In this part of work, IUF is employed as the major

benchmark and the comparison work to the proposed SAD method, because it has been

suggested to outperform many other existing similarity measures and also shares certain

concept similarity to the proposed SAD technique. The concept of IUF is briefly

described in this section.

It can be observed from Equation (A.1) and (A.2), the standard similarity

measures consider the all items equally. However, they might be further improved if the

more influenced items can be treated more importantly. The concept behinds IUF came

from the well-known information retrieval technique - Inverse Document Frequency

(IDF) (Salton, 1983) – which is commonly employed to mine the keywords from given

documents. In IDF, a word is considered less important if it occurs commonly among all

the documents. By taking this idea into the collaborative filtering, IUF suggests that the

universally rated items are less useful in capturing the user similarities than uncommon

items. For an item , the following equation can be used to measure the importance of

the item:

Page 169

| || |

where | | is the total number of users in the dataset and | | is the number of users who

have rated for item in the past. With the importance factor defined, the correlation

coefficient method can be modified by using as a weight to represent different

importance for different items. Thus, the IUF is defined as below:

,∑ , ,

∑ , ∑ ,

(A.4)

It can be easily observed that Equation (A.4) is a modification to the standard

Pearson correlation coefficient. Importantly, when comparing the items commonly rated

by and , the two users’ rating similarities towards popularly rated items are

considered insignificant (i.e. is small), in contrast, if the two users have rated

unpopular items similarly (i.e. is large), the two users will be considered as having

strong similar item preference. Therefore, is a very important factor that affects the

final results of IUF. If can be accurately computed (such that it accurately reflects the

item popularities in the entire dataset), then the logic behind IUF is indeed more

objective and appropriated than the standard similarity measures, and therefore can

greatly improve the recommenders’ recommendation quality.

A.3. PROPOSED APPROACH - SAD

In this section, the proposed Statistical Attributed Distance (SAD) method is

explained. The basic idea behinds SAD is to include the influences of attribute values

Page 170

into object correlation measurement. In essence, when comparing an attribute of two

objects, if they have the same value to in the same attribute and the value has a high

population (i.e. many objects have this value for this attribute) in the entire database,

then the similarity in terms of this attribute is considered less important. By contrary, if

the two objects have similar values to the same attribute and the value has a low

population, then the similarity in terms of this attribute is considered important. In the

case of recommender systems with user ratings as user profiles, an object is a user

represented as a vector of item ratings, each item is an attribute and the rating is the value

of the attribute. The user who have the same rating (or very similar rating) to an item and

the rating value is popularly voted to the item by many users, the likeness of the two

users’ rating to this item will not contribute much to determine the similarity between the

two users. For example, suppose that most users have voted an item with rating 7

(i.e. 7 is a popular value to this item), if two users voted this item with rating 3 which is

not popularly voted to this item by other users, the two users are considered more similar

than if they voted this item with rating 7. The concept can be further explained using

Figure A.1 and Figure A.2.

In both Figure A.1 and Figure A.2, axes x and y represent user ratings for two

items, in particular, axis x represents ratings to item and axis y represents ratings to

item . Each dot in the graph represents a user’s ratings to both items. For simplicity,

we only show the positive region (e.g. rating 4-7) and negative region (e.g., rating 1-3)

on each axis to indicate users’ preference to the items. For example, since is placed at

the middle between the positive and negative regions of both items, it indicates ’s

preferences to both of the two items are neutral. Group and are two different

sets of users grouped according to the observable similarity. The users in all rate

positively but negatively to , by contrast, the users in prefer to .

Page 171

Figure A.1 shows that the similarity between the group and user is nearly

identical to the similarity between the group and user . The difference between

and is that the users in are similar to in terms of their ratings to item ,

whereas users in are similar to in terms of their ratings to item . If the

importance of the ratings to item and are considered equally, the similarities

between and the user groups and should be similar as depicted in Figure

A.1. This is the case captured by the standard similarity measures. However, when the

concept of SAD is considered, because the popularity of positive ratings to item is

higher than to item , the user group should be considered more similar to user

than to the group (as depicted in Figure A.2).

Figure A.1. A graph for demonstrating the concept of the standard similarity

measures

Page 172

Figure A.2. A graph for demonstrating the concept of the proposed SAD technique

Based on the concept described above, similar to the factor in the IUF we

define a weight factor as below:

1| || |

where indicates the degree of the uniqueness that a particular rating value is

given to item . In the equation, | | is the total number of users who have previously

rated and | | is the number of users who rated item with a particular value .

It can be easily observed from the equation that when there are many users who rated

item with a particular , the value will be small, conversely, if only a few users

rated with a rating value , will be large. Based on the proposed weight factor

, the proposed SAD method can be formularized as below:

,∑ sv , ,

| | 

(A.5)

where

Page 173

sv , ,

, , 1            , 0 , 0

, , 1            , 0 , 0

, ,                                            

and

, ,

In Equation (A.5), sv , , denotes the weighted similarity for and ’s

ratings to item , | | is the number of items rated by both and , and the

constant 0,1 is used to adjust the importance or influence of the weight factor

in sv , , . Specifically, when equals to 0, , acts similar to

standard similarity measures, conversely, when approaches 1, users with similarity in

their uncommon tastes will be considered more important. Moreover, , is a normalised

rating based on , , which is simply the difference between ’s actual rating to and

’s average rating (i.e. ). The idea behind the normalised rating , is adopted from

Pearson correlation coefficient, the main purpose is to reduce the differences among

different users’ personal rating styles. The value of , can be either positive or negative,

when , is positive, it indicates that ’s preference to is above average, conversely,

when , is negative, ’s preference to is below average.

In order to compute the weight factor in Equation (A.5), we firstly need to

enumerate all the possible values for (in such case, needs to be a discrete variable, or

needs to be discretized first), so that we can compute the occurrences of a particular

rating value to a given item . The equation we depicted above took the simplest

approach by discretizing the user ratings into binary variables so that each rating can be

categorized into either “like” (i.e. ) or “dislike” (i.e. ). The normalised rating (i.e. , )

effectively facilitates the desired discretization process; when , 0, it indicates that

Page 174

“likes” (i.e. ), and when , 0, “dislikes” (i.e. ). Hereby, we can then

divide the set of users who previously rated (i.e. ) into two sets and ,

and they denote the set of users who like and dislike respectively. Based on the

divided user sets and , we can then compute the weight factors and

given in Equation (A.5), such that when there are many users who like (i.e.

), will be small and is large; when there are many users

who dislike (i.e. ), will be large and is small.

The use of the weight factors and in sv , , basically

follows the concept of proposed SAD described in the beginning of this section.

Specifically, when both and rated positively (i.e. , 0 and , 0 ) or

negatively (i.e. , 0 and , 0), we include the influences of rating popularity (i.e.

the weight factors and respectively) in the final score. For example, when

two users and both rated item and positively, where is a popular liked

item and is a popularly disliked item, the similarity between and ’s preferences

to will be emphasised over the preference similarity to under the concept of SAD.

That is, the value of sv , , will be larger than the value of sv , , due to

the weight factors .

In the third case of sv , , where and have completely different

preferences about (e.g. , 0 and , 0), sv , , will return a negative

value (since , , will be negative), and the weight factor will not be included in

the computation (i.e. we don’t need to emphasise on the differences between two users’

ratings to an item).

To summarise, Equation (A.5) has precisely implemented the concept of the

proposed SAD described in the beginning of this section. Note, while Equation (A.5)

Page 175

only discretizes the rating values into “like” and “dislike” (i.e. binary discretization) for

the simplicity, one can always extend the SAD concept further with more advanced

discretization methods (e.g. discretize the rating values into five levels such as “hate”,

“dislike”, “neutral”, “like”, “love”) to obtain better results.

A.4. EXPERIMENT AND EVALUATION

In this section, the experimental results we obtained from comparing the

predictive accuracy between IUF and the proposed SAD method are presented.

A.4.1. Data Acquisition

The dataset used in this experiment is obtained from MovieLens project

(http://www.movielens.org/), and it was collected through the MovieLens web site

during the seven-month period from 1997 to 1998. The dataset is cleaned up so each

user has at least 20 ratings (i.e. | | 20). The dataset contains 100,000 ratings from

900 users on 1682 movies.

From these 900 users in the dataset, 100 of them are randomly selected into the

testing user set and the rest of 800 users are then used to form the training user set. In the

testing user set, each testing user ’s ratings are divided into two parts: the training

ratings and the testing ratings , such that and . The testing

rating set contains 10 ratings that are randomly selected from (i.e. | | 10), and

the rest of the ratings are used to form the testing user ratings .

A.4.2. Evaluation Metrics

Page 176

The prediction quality of IUF and SAD is evaluated by the Mean Absolute Error

(MAE) metrics (Breese et al., 1998, Zeng et al., 2003), and it is depicted as below:

MAE∑ | , , |

| | 

(A.6)

In Equation (A.6), denotes a item involved in ’s testing rating list , and

, and , each denotes the predicted rating and actual rating that gives to

respectively.

In this experiment, the 800 users in the training user set are used to train the

prediction algorithms IUF and SAD. We then cycle through each of the 100 user in the

testing user set, and treat each of them as a target user for the prediction algorithms.

Specifically, with a given target user , Equation (A.6) is applied to compute the

prediction algorithms’ (i.e. IUF and SAD) average miss-predictions to . Then we sum

up the results for every in the testing user set, and compute the average in order to

obtain the average miss-predictions (i.e. MAE) for IUF and SAD.

A.4.3. Experimental Results

In the section, the experimental results obtained from evaluating the IUF and

SAD methods with the MAE metrics are presented. The experiment was conducted

based on training user sets of different sizes ranging from 100 to 800.

Page 177

Figure A.3. Comparison between IUF and SAD with training sets of different sizes

Our result shows that the proposed SAD method is more accurate and stable than

the IUF method. The SAD based collaborative filtering (CF) recommender is about

6.47% more accurate than the IUF based one. Moreover, it can be seen from the results,

the IUF based CF recommender can be easily influenced by the size and rating

distribution of the training dataset, whereas the SAD based CF recommender is less

susceptible to these factors. The standard deviations of the MAE results over training

sets of different sizes for IUF and SAD is 0.0167 and 0.0025 respectively, it indicates

that SAD is much more stable than IUF giving training sets of different sizes.

Page 178

Appendix B: Hybrid Parititional Clustering

Clustering techniques have been popular applied in the domain of recommender

systems for partitioning a large number of users or products into smaller groups. In

general, clustering techniques serve two purposes in recommender systems:

Improve computation efficiency – by pre-processing large numbers of

users or products into smaller groups, the computation efficiency of the

recommenders can be effectively improved as the numbers of iterations

required for traversing through each user and item are drastically reduced in

the recommendation generation process (Cöster et al., 2002, Gui-Rong et al.,

2005, Sarwar et al., 2002).

Model learning – some recommenders require models or knowledge

learning from pre-computed user or product clusters in order to generate

recommendations (Breese et al., 1998, Burke, 2002, Ghani and Fano, 2002,

Herlocker et al., 2002, Jerome and Derek, 2004) .

Even though clustering techniques have been popularly used in recommender

systems, there are only a few works that explicitly address the development of clustering

techniques in recommender systems. While the detailed use of clustering techniques has

not been a major concern in most recommender related works, many recommender

works simply adopt existing conventional clustering techniques (e.g. k-means, k-modes,

etc.) to accomplish their clustering related tasks. Although these conventional techniques

are usually well studied and easy to implement, however, many of them are not perfectly

appropriate for recommender system related applications. In this section, a novel

clustering technique, Hybrid Partitional Clustering (HPC), is proposed and explained in

Page 179

detail. The proposed HPC technique can automatically estimates the optimal number of

clusters for a given dataset, and so it can be easily adopted by recommenders as they

don’t need to manually estimate the appropriate numbers of clusters to achieve optimal

performances in their recommendation making processes. Moreover, the proposed HPC

technique allows the resulting cluster partitions to gradually update themselves when

there are updates to the datasets, and it ensures the cluster partitions are always

consistent with the underlying datasets so that the hosting recommenders are always in

the optimal states. Before the proposed HPC technique is explained, some existing and

state-of-the-art clustering techniques are briefly reviewed in Section B.1.

B.1. EXISTING CLUSTERING TECHNIQUES

Clustering is an unsupervised classification process that partitions a large set of

data or objects (or users and items in the context of recommender systems) into

homogeneous clusters. As the ‘unsupervised nature’ indicated, it is often assumed that

the clustering users have minimal information and knowledge to the data being observed.

Therefore, the major objective of clustering is to organise the mass and disorderly

objects into a set of meaningful clusters (Jain et al., 1999). Clustering plays an

outstanding role in several research fields such as scientific data exploration, information

retrieval and text mining, spatial database applications, Web analysis, customer

relationship management (CRM), marketing, medical diagnostics, computational

biology, and many others (Berkhin, 2002), therefore huge amounts of works have been

done in this area. Detailed reviews and surveys about the current state-of-the-art

clustering techniques can be found in (Berkhin, 2002, Jain et al., 1999, Pedrycz, 2005).

Clustering techniques can be broadly divided into two categories, namely,

partitional clustering and hierarchical clustering. For partitional clustering techniques,

Page 180

various partitions containing clusters are constructed, and based on some criterion the

partition that minimises (or maximises) a predefined objective function is then chosen

(Frigui and Krishnapuram, 1997, Pedrycz, 2005). However, a major shortcoming of the

partitional clustering is that the number of resulting clusters (i.e. ) has to be specified in

advance, and it is difficult for users to supply the exact value of manually when their

knowledge to the data is limited. Moreover, some partitional clustering techniques such

as k-means and k-modes are prone to local optimum and their clustering results are

sensitive to initial locations of the cluster centres (i.e. these techniques often randomly

select points in the initialisation stage, and iteratively adjust them to the correct cluster

centre locations to form clusters). Conversely, hierarchical clustering techniques create a

hierarchical decomposition of dataset and it is often represented in a form of dendrogram.

A partition in hierarchical clustering can be obtained by cutting the dendrogram at some

desired level, and therefore it is not required to specify the number of output clusters in

advance. Notwithstanding hierarchical clustering provides better analytic features than

partitional clustering (as data can be visualised in a dendrogram), it generally does not

scale well for large datasets. In addition, in classical hierarchical clustering (e.g.

agglomerative and divisive based hierarchical clustering) objects that are committed to a

cluster in the early stages cannot move to another cluster. In other words, once a cluster

is split or two clusters are merged, the split objects will never come together in one

cluster or the merged objects will always stay in the same cluster, no matter whether the

splitting or the merge is a right action or not. It is shown in (Pelleg and Moore, 2000, Xu,

2005), some previous splitting or merging actions in hierarchical clustering maybe not

right and some split and merged objects may need to be rearranged in latter actions. This

particular issue is the major cause of inaccuracy in hierarchical clustering, especially for

large datasets.

Page 181

Besides these two classical techniques, there are also some extensions and

advanced clustering techniques designed to address the limitations of the classical

clustering techniques. The X-means suggested by Pelleg and Moore (2000) is one of the

most popular extensions to classical k-means. X-means saves users from specify the

exact (i.e. number of resulting clusters); instead, users only need to specify a possible

range of , and the X-means will return the optimal partition within the specified range.

Likas et al. (2003) tries to produce the optimal partition by using an incremental

technique that dynamically add one cluster centre at a time through a deterministic

global search procedure from suitable initial positions. Pelleg and Moore (1999) utilise

the kd-tree data structure and the geometric reasoning techniques to estimate the initial

locations of the cluster centroids. In contrast to Pelleg and Moore (1999)’s work, Al-

Daoud (2005) proposed a less sophisticated centroid initialisation method based on

finding a set of medians extracted from a data dimension with maximum variance.

The proposed HPC technique addresses not only the limitations of both the

standard partitional and hierarchical clustering techniques, but it also provides some

advantages over other advanced clustering techniques. A general overview of HPC’s

algorithmic concept and some comparisons between HPC and the existing clustering

techniques described above are provided in Section B.2.

B.2. GENERAL OVERVIEW

The proposed HPC technique consists of three consecutive phases: initial

centroids estimation, partitional clustering and hybrid partition adjustment and

optimisation (as depicted in Figure B.1). In the first phase, for a given dataset the most

possible number of clusters and the possible centroids of the potential clusters are

estimated with a novel centroid estimation technique. In the second phase, the estimated

Page 182

centroids are utilised to initialise a standard partitional clustering technique (e.g. k-

means), and then the initial cluster partition can be obtained by executing the selected

partitional clustering technique. In the final phase, a incremental clustering algorithm,

Hybrid Hierarchical Clustering Algorithm (HHCA), proposed by (Xu, 2005) is

employed to further optimise the initial cluster partition resulted from the second phase

based on a predefined objective function.

Figure B.1. The three major consecutive phases of the proposed HPC technique

One of the advantages provided by the HPC technique is that neither the number

nor the range of the possible clusters needs to be specified in advance. Therefore, the

HPC technique provides better usability than standard partitional clustering techniques

Page 183

such as k-means, k-modes, etc. or even more advanced techniques like X-means, which

require the range of resulting clusters to be pre-specified.

Partitional based clustering techniques such as k-means and k-modes usually

provide only local optimal clustering solutions. It is mainly because their clustering

results are strongly dependent on the initial centroids selections which are often based on

randomisation (Jain et al., 1999). Hence, when conducting multiple trials in one dataset,

partitional based techniques (e.g. k-means and k-modes) usually produce clustering

results with different quality. Specifically, when they are initialised with centroids that

are closer to the true centroids locations, the efficiency and the resulted clustering

qualities can be greatly improved, conversely, poor chosen centroids for the initialisation

might result in poor performance and clustering results. Standard techniques based on

randomised centroids initialisation (i.e. k-means, k-modes etc.) usually need to be

executed in numerous runs (with different centroids initialisations) in order to determine

which clustering results are closer to the optimal solution. Obviously, such solution is

impractical, inefficient and error-prone when the target dataset is large. As mentioned

before, some existing centroids location estimation techniques have been proposed to

improve the performance of the partitional based techniques (e.g. (Al-Daoud, 2005,

Pelleg and Moore, 1999)) , and they all reported to have achieved certain amount of

improvements over standard techniques in their experiments. However, to the best of

our knowledge, none of them can both automatically estimate the number of centroids

and their corresponding locations for the given dataset, and many of them can only

estimate the centroid locations with the number of centroids been manually specified in

advance. In contrast, the proposed centroids estimation technique used in the first phase

of the HPC can estimate both the number of centroids and their locations for any given

datasets. Hence it not only provided a better usability than other techniques but also

Page 184

enables the partitional clustering technique used in the second phase to perform more

efficiently and result better quality for the initial clustering partition.

In the final phase of HPC, the initial clustering partition resulted from the second

phase is further optimised in accordance to a predefined objective function (see Section

B.3). The purpose of the objective function is to allow users to specify the desired cluster

granularity while not interfere the overall clustering quality. Specifically, depending on

different usages, users can specify if they need the clustering results to have a large

numbers of clusters with higher densities in each of the clusters or a smaller numbers of

clusters and each with lower densities. This design provides better usability than both

partitional and hierarchical based techniques, because it enables the users a certain

flexibility to control their desired clustering results and allows them to have minimal

knowledge to the target datasets (i.e. do not need to know the size or the density of the

dataset). In order to optimise the initial clustering partition from the second phase, the

employed HHCA iteratively merges and splits the clusters in the partition until the

objective function is maximised. In particular, unlike standard hierarchical clustering

techniques where clusters can only be merged or split but not both, HHCA allows

clusters to be split or merged in every partition updates. Hence given two objects that

have been divided into two different clusters, they might be merged into one cluster in

latter update iterations. This feature allows HHCA to produce better clustering results

than other hierarchical based techniques and also be able to cope with frequent dataset

updates (i.e. when the objects are added, removed or modified in the datasets, the

corresponding clustering partitions can be efficiently adjusted based on the changes).

The details of the HPC and the three phases are to be detailed in the following

sections. In Section B.3, the objective function we employed to evaluate the quality of

Page 185

resulting partition is summarised. Section B.4, B.5 and B.6 explain the three consecutive

phases of the proposed HPC technique respectively.

B.3. OBJECTIVE FUNCTION

As in the notion of the partitional clustering, a clustering problem can be

considered as an optimisation problem to a predetermined objective function. In this

work, the objective function defined in (Xu, 2005) is employed. By solving the

maximum of objective function, the resulting partition will have maximum intra-cluster

similarity and maximum inter-cluster distance. In other words, it is expected that the

objects within a cluster are as close as possible and the objects in different clusters are as

far as possible.

Let , , … , be a set of given data objects, where each data point

can be represented as a p-dimensional vector in a vector space. For a given , we

assume , , … , be a partition over the dataset where   for all

and . Moreover, the cluster centroid (i.e. cluster median or central

id) for is denoted as , , , , … , , , where , indicates the dimension

of ’s cluster centroid. Note, because HPC is a very general technique and can be

applied to many different applications (not just for recommender systems), therefore, we

will use some new notations in this section that are less specific to recommender systems

(i.e. different notations from those employed in previous chapters). For example, an

object can be either a user or an item in recommender systems depending on the target

recommender types (e.g. collaborative filtering vs. item-to-item collaborative filtering).

However, for the understandability, readers can assume that each object

corresponds to a user in a recommender system, the object attributes are the user

Page 186

ratings, and the goal is to divide the entire user set into user clusters (i.e.

corresponds to ).

Before defining the cluster intra-similarity and inter-distance, it is important to

specify how the similarity and distance between two objects (or data points) are

measured. In this work the two commonly used measurements, cosine similarity and

Euclidean distance measure, are chosen to measure the similarity of objects:

,  

(B.1)

,  

(B.2)

where , are two objects.

When and   are considered as two users in a recommender system, their

cosine similarity and Euclidean distance can be computed by Equation (B.1) and (B.2)

respectively. Moreover, depending on the target dataset and application, the cosine

similarity and Euclidean distance measures can be replaced with other similarity and

distance measures such as those described in Appendix A.

The intra-similarity of a cluster is simply the average of the similarities

between all the objects within and the cluster centroid . Specifically:

intra_sim∑ ,

| | 

(B.3)

Page 187

Based on Equation (B.3) we can then further measure the average cluster intra-

similarity of a given partition (or a clustering result) :

p_intra_sim∑ intra_sim

| | 

(B.4)

While cluster intra-similarity measure can be used to determine the cluster

densities (i.e. whether the objects within in a cluster are closed to each other), we also

need to be able to measure the distance between two different clusters (i.e. whether the

objects in different clusters are far way from each other). In this work, the distance

between two clusters is measured by calculating the distance between their centroids:

cluster_dist , dist ,  

(B.5)

Based on Equation (B.5), we can then evaluate the overall cluster inter-distances

of a given partition by simply averaging the distances of all the possible cluster pairs

from the partition:

p_inter_dist∑ cluster_dist ,, ,

| | | | 1 

(B.6)

Finally, by combining Equations (B.4) and (B.6), the objective function (i.e.

quality of a given partition) is given by:

Page 188

p_qual p_inter_dist 1 p_intra_sim  

(B.7)

where 0 1 is used to adjust the weights of the cluster inter-distance and intra-

similarity in the final partition quality scores.

By observing closely the cluster inter-distance and intra-similarity measures

defined in Equations (B.4) and (B.6), it can be seen that they have different reflections

on the partition granularities. Specifically, the cluster intra-similarity measure trends to

give higher scores to partitions with large number of small clusters, because small

clusters usually have higher densities (i.e. cluster intra-similarities) than large clusters. In

contrast, the cluster inter-distance measure trends to give higher scores to partitions with

small number of large clusters, because the centroids of large clusters are usually more

distant from each other than centroids of small clusters. Hence, the control parameter

used in Equation (B.7) can be used to adjust the desired partition granularities. When

is set to values closed to 1, the cluster inter-distance is considered more important than

the cluster intra-similarity, thus, Equation (B.7) will give higher scores to partitions with

small number of large clusters (i.e. partitions with low granularities). Similarly, when

is set to values closed to 0, the cluster intra-similarity will receive higher weight, and

hence Equation (B.7) will give higher scores to partitions with large number of small

clusters (i.e. partitions with high granularities). Thus, by adjusting the control parameter

, users can easily and effectively adjust the desired granularities of the cluster partitions

generated by the proposed HPC technique.

B.4. CLUSTER CENTRE ESTIMATION

Page 189

The proposed cluster centroid estimation technique is intuitive, effective and

considerably efficient, and provides the following three features:

Estimation of the possible number of potential clusters.

Estimation of the centroid locations of the potential clusters.

Outlier detection.

B.4.1. Cluster Centroid

Before going into the details of the proposed centroid estimation technique, it is

important to understand the basic nature of a cluster centroid.

In k-means, a cluster centroid is the average of all the data points in a cluster. In

other words, its coordinates are the arithmetic mean for each dimension separately over

all the points in the cluster. On the other hand, the cluster centroid in k-modes is the

median data point in the cluster. A more descriptive explanation to the concept of cluster

centroid is given in Fuzzy C-Means (FCM) (Berkhin, 2002, Pedrycz, 2005):

,  

(B.8)

where is the centroid of cluster and is a membership function measuring the

likelihood that the object belongs to the cluster . The C-means clustering method is

to find cluster which maximise Equation (B.8) . Equation (B.8) indicates the following

features:

Page 190

A cluster centroid is an object (can be either a virtual or actual object or data

point) within a cluster, such that the distances between it and all other

objects in the cluster are averagely shortest.

The distances between the cluster centroid and all other objects outside the

cluster are insignificant (i.e. they were filtered out by given very small

values of ).

B.4.2. Single Cluster Centroid Estimation

For the simplest case where the dataset containing only a single cluster (i.e.

| | 1), we can find the arithmetic mean or median of the objects in the cluster as the

centroid, and it is the commonly used approach in k-means and k-modes. In our case,

however, it is required to find the cluster centroids based on only the distances among

the data points, and there are several reasons for it:

Standard arithmetic mean or median computes only centre locations for

given clusters, however, the computed centroid results do not contain

information about the cluster densities. However, in the proposed centroid

estimation algorithm, it is required to compare centroids of multiple clusters

based on their cluster densities. Hence, standard arithmetic mean or median

for computing cluster centroids does not suit the proposed centroid

estimation algorithm.

Standard arithmetic mean or median is only applicable to objects in standard

vector space. However, they might not be applicable to other complex

objects (e.g. objects with categorical attributes).

Standard arithmetic mean or median might be inconsistent with the

similarity measures employed. As a cluster centre needs to be similarly

Page 191

distant to all of the objects within a cluster, therefore, it should be dependent

on the similarity (or distance) measure employed. However, standard

arithmetic mean or median is independent of the similarity measure

employed. Hence, when advanced similarity measures such as Pearson

correlation coefficient, IUF, SAD, etc. are employed, the centroids

computed by the standard arithmetic mean or median might not be the

desired ones.

In order to find the cluster centroid based on only object distances, we define

as the weight of the object , and specifically:

,\  

(B.9)

Then, the possible centroid of (note, we assume there is only one cluster in , so

all objects in is contained in that cluster) can be estimated by:

  arg max    

(B.10)

is the object that is close to all other objects in the cluster averagely. It can be

observed that Equation (B.9) has a strong connection with Equation (B.8). In Equation

(B.9) it is assumed that all objects in the cluster are possible centroids, and the

membership function in Equation (B.8) is replaced by , .

Page 192

As similar to Equation (B.8), the larger value of indicates the larger

possibility that is the cluster centroids. Hence, the cluster centroids for a single cluster

dataset can be obtained by resolving Equation (B.10). Figure B.2 depicts a possible data

distribution of a single cluster, and the objects with their corresponding weight values

(computed by Equation (B.9)) are depicted in Figure B.3.

Figure B.2. A possible dataset with a single cluster

Figure B.3. An example of centroid estimation based on Equation (B.10)

Page 193

B.4.3. Multiple Cluster Centroids Estimation

For more realistic datasets containing multiple clusters and outliers (e.g. Figure

B.4), the technique described in Section B.4.2 is insufficient. Figure B.5 depicted the

result when we use equation (B.9) to compute the weights for the objects. From Figure

B.4 and Figure B.5, it can be observed that even though cluster A contains more objects

and is more crowded than cluster B, however, almost all of the objects in cluster B have

higher weight values (i.e. ) than the objects in cluster A. Also, the object with the

highest weight value in Cluster A is no longer in the cluster centre, and instead it is now

at the edge of cluster A between cluster A and B. Moreover, even though cluster C and

A contain similar numbers of data points and at similar positions (i.e. both locate in the

corners of the plane), most points in cluster A have higher weight values than cluster C

because cluster A is more condensed (i.e. with higher density).

Figure B.4. A possible dataset containing multiple clusters

Page 194

Figure B.5. Centroids estimation for the complex dataset with multiple clusters based

on Equation (B.10)

Figure B.6. An example of virtual boundaries for each of the clusters in the dataset

To summaries, the reasons that why Equation (B.9) failed to produce higher

values for possible cluster centroids of the clusters in the dataset are:

It trends to produce higher weight values for the objects at the centre of the

dataset instead of the centres of the clusters.

Because the size and density are different for different clusters, the weights

of the data points for one cluster are not comparable to other clusters. That is,

the cluster centre of a sparse cluster might have a smaller weight value than

most of the data points in a condensed cluster.

In order to estimate the cluster centroids for dataset consisting of multiple

clusters with different densities, we need to revise the weight computation algorithm so

Page 195

that the weight value of an object won’t be influenced by objects in other clusters. That

is, when computing the weight for an object, we need to estimate a cluster boundary

(such as the circles plotted in Figure B.6) so that any points outside the boundary is

considered inexistent. In this work, we proposed a simple technique using the average

shortest distance of the dataset to estimate the boundary length for the dataset as

described in Equation (B.10):

∑

| | 

(B.11)

where is the distance from to its nearest neighbour object, that is:

min,

,

Based on the computed by Equation (B.11) we can then find the

neighbour objects within the boundary of a given object . The set of neighbour

objects of denoted as , is estimated by:

| , ,

only contains the objects that are most likely in the same cluster with .

Therefore, when adopting this boundary constraint to Equation (B.9), the unnecessary

influences from the objects of other clusters are effectively reduced. Specifically,

Equation (B.9) can be modified to:

,,  

(B.12)

Page 196

With Equation (B.12) the proposed cluster centroid estimation algorithm, thus,

can be described as below:

Algorithm B.1. _

Input is the object set to which the centroids are to be estimated.

Output is the set of estimated cluster centroids of

1) SET as the set of the estimated centroids and it is initially empty.

2) Find the most possible centroid for the current dataset X based on Equation

(B.12) and Equation (B.10). Specifically:

arg max  

3) SET . That is, add the most possible centroid of the current dataset to

.

4) SET \ . Remove the centroid from the dataset.

5) SET \ . Remove all ’s neighbour objects from the dataset.

6) IF | | 0

7) THEN return Z as the set of estimated cluster centroids.

8) ELSE go to step 2.

The idea behinds the proposed method is quite intuitive. Firstly, in line (2) of the

algorithm, the object in the centre of the most crowded object group is considered as the

most possible centroid . Next, the estimated centroid is recorded (i.e. line (3)) and

removed from the dataset (line (4)). Then all neighbour objects of are also removed

from , because they have higher possibility to be in ‘s cluster than to be the centroids

of other clusters (line (5)). The procedure from line (2) to line (7) of the algorithm is

Page 197

repeated until no more possible centroids can be found in (i.e. when becomes

empty).

As example is given in Figure B.7 which illustrates the process of estimating the

cluster centroids using Algorithm B.1. In Figure B.7(a) the centroid in cluster A has the

highest weight than the centroids in cluster B and C as cluster A has the highest density

(i.e. all objects in the cluster are close to each other). After the most possible cluster

centroid has been detected as shown in Figure B.7(a), the detected centroid and its

surrounding neighbours are removed from the dataset in order to allow the centroids of

other clusters to be detected in following rounds. Figure B.7(b) shows the resulting

dataset after the removal of detected centroid and its neighbour objects in Figure B.7(a).

Similarly, Figure B.7(c) shows the resulting dataset after the second most possible

centroid and its neighbour objects are removed from the dataset in Figure B.7(b). It is

worth noting that the last few estimated centroids are very likely to be outliers (see

Figure B.7(d)) , hence the proposed technique can also be used for outlier detection.

Figure B.7. An example of cluster centroids estimation process

B.5. PARTITIONAL CLUSTERING

Page 198

As mentioned earlier, the centroids detected with the centroid estimation

technique described in Section B.4 are not perfectly accurate, they are mainly designed

as the initial centroids for using partitional clustering techniques to complete clustering

tasks. Algorithm B.2 given below is a modified k-means method, which uses the

estimated centroids as the initial centroids of possible clusters.

Algorithm B.2 . _

Input , … , is the set of estimated centroids returned from

Algorithm B.1 (i.e. _ ).

Output is the resulting cluster partition for the dataset

1) SET , … , as the initial partition consisting | | empty clusters,

specifically, :  .

2) Associate each cluster with a corresponding centroid from , so that

denotes the centroid of .

3) Assign the objects in to their nearest clusters, such that:

:  | , min ,

4) Update the cluster centroids by computing the arithmetic means of the clusters, and

let be the new set of cluster centroids.

5) IF

6) THEN return as the resulted partition.

7) ELSE SET and go to step 3.

By initialising the k-means with the estimated centroids, the possibility of

obtaining local optimal partition results is reduced. In order to demonstrate the

Page 199

effectiveness of the estimated centroids to the k-means method, a simple experiment is

conducted to evaluate the computation efficiency and the clustering quality (calculated

based on Equation (B.7)) by comparing with the following three techniques:

The standard k-means technique with a randomly generated .

The modified k-means method proposed in this section, i.e. initialising the

standard k-means technique with the estimated centroids as described in

Algorithm B.2.

The standard k-means technique with the estimated . That is, instead of

initialising the k-means with the estimated centroids, we only use the

number of the estimated centroids ( ) as the initialisation parameter.

The datasets employed in the experiments are sets of randomly generated two

dimensional vectors with different sizes and densities. The experimental results for

evaluating the clustering quality and computation efficiency are depicted in Figure B.8.

and Figure B.9. respectively. Note, in the computation efficiency experiments (i.e.

Figure B.9. ), the computation efficiency of the standard k-means is not included in the

comparison. It is because the major purpose of the experiments is to test whether the

predicted centroids are relatively accurate so that the iterative centroids refinement

process in the standard k-means can be effectively reduced (and therefore results in

better computation efficiency). The computation efficiency of the standard k-means

(both and centroids are randomly generated) is unnecessary because it is difficult to

determine whether the computation efficiency of the standard k-means comes from the

randomly generated or centroid locations.

It is shown in Figure B.8. that the partition quality (measured based on

Equation (B.7)) achieved by initialising the standard k-means with the estimated is

improved comparing to the standard k-means with a random . This result demonstrated

Page 200

that the number of clusters estimated with the proposed technique is relatively close to

the true number of clusters in a dataset. Moreover, it is also shown in Figure B.8. that by

including both estimated and centroid locations in the k-means, the best clustering

results are achieved, hence it can be further concluded that the centroid location

estimated by the proposed method is accurate as well.

In Figure B.9. , it is shown that by including the estimated centroid locations

into the k-means the computation efficiency can be achieved almost twice as efficient as

the k-means with randomly selected centroids. This improvement suggests that the

estimated initial cluster centroids are close to the correct locations, so that the amount of

time required to find the correct centroids locations is greatly reduced.

To summarise, based on the results obtained from this simple experiment, it can

be concluded that the proposed centroids estimation technique is accurate and its

application to the standard partitional clustering algorithms (i.e. k-means) is beneficial in

both computation efficiency and clustering partition quality.

Figure B.8. Partition quality comparison with different k-means settings

Page 201

Figure B.9. Computation time comparison with different k-means settings

B.6. PARTITION ADJUSTMENT AND OPTIMISATION

In the third phase of the proposed HPC approach, the HHCA method (Xu, 2005)

is employed to fine-tune the clustering partitions so the objective function (B.7) can be

satisfied. The HHCA method is different from the standard hierarchical clustering

methods, it allows the previously committed clusters to be revised in both divisive and

agglomerative manners. With HHCA, a partition is iteratively fine-tuned by comparing

the current partition with new partitions generated by a divisive strategy and an

agglomerative strategy. The detail of HHCA is described below:

Algorithm B.3

Input , … , is the target partition generated from Algorithm B.2 (i.e.

_ ) for the further adjustment and optimisation.

Output is the resulting cluster partition for the dataset

1) SET , … , as the initial partition consisting | | empty clusters,

specifically, :  .

2) Customise the objective function (Equation (B.7)) by setting the partition

granularity control parameter (as described in Section B.3).

3) Find a cluster with minimal intra-similarity:

Page 202

argmin  intra_sim

4) Find the two most dissimilar objects and in :

 ,   argmax,

 dist ,

5) Dividing into and based on and , such that:

| , ,

| , ,

6) Create a new partition by removing from and adding the two clusters

and created in step 5:

\ ,

7) Find the two most similar clusters and in :

, argmin, ,  

cluster_dist ,

8) Create a new partition by removing and from and adding in the union

of and :

\ ,

9) Evaluate the qualities of the three partitions , and based on Equation (B.7).

10) IF p_qual p_qual and p_qual p_qual

11) THEN SET and go to step 3.

12) ELSE IF p_qual p_qual and p_qual p_qual

13) THEN SET and go to step 3.

14) ELSE return as the resulted partition.

Because the main function of the HHCA algorithm is to optimise an already

existed partition, it can also be used to update existing cluster partitions. When a cluster

partition is constructed from a dataset, it is possible that the latter updates (i.e. adding,

Page 203

removing and modifying objects) to the dataset might reduce the quality of the cluster

partition. Hence, the proposed HHCA technique can be employed to optimise the

partition after the dataset updates. Since the HHCA technique updates existing partitions

incrementally along with the dataset updates (i.e. do not need to execute the entire

clustering process from scratch), it ensures competitive computation efficiency for the

partition update process.

B.7. EXPERIMENT AND EVALUATION

In this section, the experimental results we obtained from evaluating the

efficiency and effectiveness of the proposed HPC technique are presented.

B.7.1. Data Acquisition

The experiments described in this section were conducted using web server logs

of individual browsing records for users at the msnbc.com site. The server-log files have

been converted into a set of browsing sequences, one sequence for each user session, and

the sequence is represented as an ordered list of category indicators. An example of the

user sequences is given in Table B.1.

Table B.1. An example experimental dataset

User Browsing Sequence

1 FRONT PAGE, NEWS, TRAVEL, TRAVEL

2 NEWS, NEWS, NEWS, NEWS, NEWS

3 FRONT PAGE, NEWS, FRONT PAGE, NEWS

Page 204

4 FRONT PAGE, SPORTS, NEWS, NEWS

5 WEATHER

The clustering task for this dataset is to group these browsing sessions based, so

we can observe and analyse the different types of browsing behaviours. There are

originally 989818 user browsing sequences in the dataset, and each user has visited

around 5.7 pages averagely. In our experiment, we removed from the log files the users

who visited less than 5 pages or more than 10 pages, so the dataset contains only 10000

users after pruning. It might be noticed the user browsing sequences (data points) belong

to categorical data, therefore, Equation (B.1) and (B.2) are no longer suitable for

measuring the similarity or distance of the data points. In the experiment we used the

percent disagreement to measure the distance between two data points:

,∑ ,

where is the vector dimension of the data points, denotes the dimension value of

, and

,0,

1,

B.7.2. Evaluation Metrics

One of the most obvious ways to evaluate the effectiveness of the clustering

techniques is to examine the quality of the resulting cluster partitions. In the experiment,

the partition quality measure depicted in Equation (B.7) is used to evaluate the cluster

partitions produced by the experimental clustering algorithms.

Page 205

B.7.3. Experimental Results

In this experiment, we compared the performances of three different clustering

algorithms. The first one is the standard HHCA method as described in Algorithm B.3

which generally performs better than classical agglomerative single-link clustering

algorithm (ASLCA) (Xu, 2005), therefore, we decided to use it as our baseline for

evaluation. The standard k-means method is also employed for the comparison where the

k is determined by the result partition generated by the HHCA (however, the centroid

locations are randomly chosen). The last algorithm to be included in this experiment is

therefore the proposed Hybrid Partitional Clustering (HPC) method.

Figure B.10 and Figure B.11 represent the average intra cluster similarity and

inter cluster distance of the resulting partitions obtained from applying the three different

methods over dataset with different sizes. Figure B.12 depicts the resulting partition

qualities (i.e. the combination of Figure B.10 and Figure B.11), where the   parameter

(see Section B.3) is set to 0.5 so the inter cluster distance and intra cluster similarity are

considered equally in the evaluation. From the results, we can see that overall the

proposed hybrid partitional clustering methods outperforms other two methods and the

k-means method is relatively unstable and therefore results in the poorest quality

partitions.

Page 206

Figure B.10. Intra-cluster similarity of the resulting cluster partitions

Figure B.11. Inter-cluster distance of the resulting cluster partitions

Figure B.12. overall quality of the resulting cluster partitions

Page 207

Appendix C: Relative Distance Filtering

As mentioned earlier in this chapter, one possible way to ensure the scalability

and efficiency of recommenders is to improve the scalability and efficiency of their

neighbourhood formation process. More precisely, given a target user or item, the goal is

to improve the computation efficiency for finding a subset of users or items from a large

dataset with high similarities (or short distances) to them.

In order to improve the efficiency of the neighbourhood formation process, many

recommenders adopted clustering techniques to reduce neighbourhood searching spaces.

For example, the proposed HPC technique described in Section Appendix B is a

relatively efficient and accurate clustering algorithm specially designed for recommender

systems. Despite their popularity, clustering based recommenders are usually weak in

coping with frequent dataset changes and updates (see Appendix B for more details),

because it is computationally expensive to rebuild a new partition from scratch when the

underlying dataset is updated. The proposed HPC technique is an alternative way to

allow incremental partition updates (i.e. running the update process from existing

partitions), it is still expensive to update cluster partitions for every small dataset changes.

Hence, many works have suggested that the partition update process can be run in offline

with less frequency (e.g. every one or two days), however, such compensation might

result in poor recommendation quality. Clustering is a way to construct neighbourhoods.

However, for large datasets, the size of each cluster may be still too big to accurately

allocate the most similar objects for a given object. That means, if the size of the cluster

is large, we will need to further retrieve a subset of objects from the cluster that have

high similarity to the target object. In this section, a novel neighbourhood estimation

method called ‘relative distance filtering’ (RDF) is presented. The basic idea of the RDF

Page 208

method is to pre-compute a small set of relative distances between objects offline, and

then using the pre-computed distance to eliminate most unnecessary similarity

comparison between objects when forming the neighbourhood for a given object. The

proposed RDF method is capable of dynamic handling frequent data update; whenever

new objects are added to the dataset, or existing objects in the dataset are deleted or

modified, the pre-computed structure cache can also be efficiently updated. Moreover,

the proposed RDF method can be used to improve clustering efficiency. For example in

the standard k-means technique, the RDF method can be used to efficiently reallocate

objects in the dataset to their closest centroids for every centroids update iterations (e.g.

step 3 of the Algorithm B.3). Also, the efficiency of the proposed centroid estimation

technique described in Section B.4 can be effectively improved by using the RDF

method to retrieve the nearest neighbours for any given object (i.e. the required

computation time for Equation (B.11) and step 5 of Algorithm can be reduced).

The most common approach nowadays for improving the efficiency of the

nearest neighbour search tasks is by tree structure based indexing techniques such as R-

Tree, kd-Tree, etc. However, these index based techniques are usually inaccurate and

memory inefficient when the target dataset consists of only high dimensional objects.

Unfortunately, the user profiles or item contents in recommender systems are usually

having very high dimensionalities (e.g. a user might be represented by a vector with the

number of dimensions equals to the numbers of books in the dataset). In contrast to these

techniques, the proposed RDF method is both memory and computation efficient even

when the target objects are in very high dimensions. In our experiments, by adopting the

proposed RDF technique to the standard recommender systems, both the computation

efficiency and recommendation quality of the recommenders are improved.

Page 209

C.1. PROPOSED APPROACH

Depending on the different types of the recommender systems, the target objects

for the neighbourhood formation process may have different types. For example, in

standard collaborative filtering based recommender systems, the goal is to locate users

with similar tastes to the target user. In contrast, the target objects in the content-based

recommenders are basically items represented by keywords vectors. While the proposed

RDF technique can be used for searching objects with various types, however, for the

simplicity of the discussion, it is assumed that the target object type is ‘user profile’ and

the goal is to find similar users for any given target user.

Forming neighbourhood for a given user with standard ‘best-n-

neighbours’ technique involves computing the distances between and all other users

and selecting the top neighbours with shortest distances (or highest similarities) to .

However, unless the distances between all users can be pre-computed offline or the

number of users in the dataset is small, forming neighbourhood dynamically can be

expensive.

Clearly, for the standard neighbourhood formation approach described above,

there is a significant amount of overhead in computing distances for users that are

obviously far away (i.e. dissimilar users). The performance of the neighbourhood

formation can be drastically improved if we exclude most of these very dissimilar users

from the detailed distance computation. In the proposed RDF technique, this exclusion

or filtering process is achieved with a simple geometrical implication: If two points are

very close to each other in a space, then their distances to a given randomly selected

point in the space should be similar.

Note, the geometrical implication described above is unidirectional, that is, it

does not imply that if the two points’ distances to a given randomly selected point are

Page 210

similar then they will be in similar position. For example, it is shown in Figure C.1 that

points , and are closed to each therefore their distances to point are similar (i.e.

). However, even though ’s distance to is also similar to the distances

from a, and to , it can be easily observed from the figure that is distant from ,

and .

Figure C.1. A simple example of the suggested geometrical implication

In addition, the proposed geometrical implication can also be supported by the

theorem of inverse triangle inequality (Saitoh, 2003). Specifically, giving any three

objects , b and c in a space, the theorem states that the distance between any two of

these objects (e.g. ) is larger than (or equal to) the difference between these two

objects’ distances to the third objet (i.e. ). Formally:

Based on the above equation, assume  and are closed to each other (e.g. b 0), it is

expected that should be closed to 0 as well (i.e. ). Hence, the validity

of the suggested geometrical implication is confirmed.

Page 211

Figure C.2. An example of projected user set

In order to demonstrate how the suggested geometrical implication can be

utilised by the proposed RDF technique to facilitate the neighbourhood formation

process, a small dataset of 1000 synthesised user profiles is used as a running example

for explaining the concept of the proposed RDF technique. In Figure C.2, the user

profiles in the dataset are projected onto a two-dimensional plane where each user profile

is depicted as a dot on the plane. In the figure, is the target user, and the dots

embraced by small circles are the top 15 neighbours of . The RDF technique starts by

randomly selecting a reference user in the user set, and then ’s distances to all

other users are computed and sorted.

Based on the suggested geometrical implication, it is easy to observe that all ’s

neighbours have similar distances to . Hence, in the process of forming ’s

neighbourhood, we only need to compute distances between and the users in set ,

which is defined as:

Page 212

| , , ,  

(C.1)

where , is the distance between the two user profiles and , and

it can be computed by Equation (A.2) or any other distance or similarity measures (such

as those described in Appendix A).

In Equation (C.1), , , is the difference between the

distances from to and to . According to Modus tolens inference rule, i.e. if

the consequent of an implication is false, the antecedent of the implication must be false,

from the geometrical implication mentioned above, if , , is

large, then and are not close to each other. A distance threshold is used to

determine if , , is small or large. If ,

, | is larger than   , the user can be excluded from ’s

neighbourhood. In our experiment, is set to the one tenth of the distance between the

reference user and its furthest neighbour .

To further improve the computation efficiency, we can select more reference

users (for example and ) into the estimation process to obtain more estimated

searching spaces (i.e. and ). With multiple estimated searching spaces, the final

estimated searching space can be drastically reduced by intersecting these spaces

(i.e.  ). It can be observed in Figure C.3 that the intersected searching space

(i.e. the two areas indicated as ‘estimated neighbourhood search space’ in Figure C.3) is

much smaller than the entire set, and most importantly, it covers ‘s most nearby users.

Because only the users in the intersection area need to be checked in order to determine

‘s final neighbourhood, the actual I/O (i.e. retrieving user profiles from the databases)

Page 213

and distance computations are therefore reduced to within only the intersected space,

hence the efficiency is greatly improved.

Figure C.3. Estimated searching space with three reference users

In order to optimise the computation efficiency with multiple reference users, the

final estimated searching space (i.e. ) should be as small as possible for any

given target user. In order to achieve it, the distances between the reference users need to

be as far as possible. It is because if the reference users are close to each other, the ring

borders of their search spaces will result in large overlap (since they all have similar

centres and radiuses). Moreover, the number of reference users should be kept small (we

only use 3 reference users for all our experiments), because when the number of

reference users increases, the computation time required for the offline reference user

initialisation and the memory required for caching the sorted distances increase too.

In our implementation, the reference users are initialised with a simple two-pass

technique. The first reference user is chosen randomly, and we compute its distances

to all other users in . Next, with the computed distances we can obtain the second

reference user such that:

argmax    ,

Page 214

Finally, we again find the furthest neighbour for and such that:

argmax    , ,

, and set as the third reference user. With this method, it is ensured that the

initialisation process is kept simple and efficient, and the result reference users are also

very distant from each other.

C.2. PROPOSED IMPLEMENTATION

This section describes in detail the implementation of the proposed RDF

technique discussed in Section C.1. With the proposed implementation, the power of

RDF is maximised.

First of all, it is important to note that the distances between users and reference

users are not meant to be computed online, because the computation efficiency of this

process is more expensive than the one by one search. Instead, these distances are pre-

computed, structured and indexed offline into a data structure called RDF searching

cache, and the searching cache will be loaded into the memory in the initialisation stage

of the online recommendation process. This pre-computed searching cache is shared by

all neighbourhood formation processes. The detailed structure is depicted in Figure C.4.

In the searching cache, each user profile is associated with a data structure called

‘user node’. For any user , denotes ’s user node. A user node basically stores

two types of information for a user:

User ID: Instead of fitting the entire user profiles into memory, only the

user id is required to be stored in the cache. The user ids are used to identify

and retrieve the actual user profiles in the database.

Page 215

Distances to the reference users: The distances from the user node’s

corresponding user to the reference users are stored in a vector. In our

implementation, we have only reference users , and , and therefore

the distance vector for user node is

, , , , , . We denote the distance vector

of as , , where corresponds to , ,

corresponds to , and corresponds to , .

In order to efficiently retrieve the estimated searching space as described in

Equation (C.1), a binary tree structure is used to index and sort the user nodes. The index

keys used for each user node are the distance between the users and the reference users,

that is, the index keys for are , and . With the three different index keys, the

user nodes can be efficiently sorted with different index key settings, that is, the user

nodes can be sorted by any one of the three index keys.

Because the user nodes are stored in this binary tree structure, the computation

efficiency for Equation (C.1) is optimised to , where | | . Note, this

estimated user space retrieval process is very efficient, not only because the whole

computation can be done within a small amount of memory (thus, no database I/O is

required), it is also because each index key lookup involves only a comparison of float

values (i.e. the distances to the reference users). Finally, because distances between the

target users and the reference users are needed during the neighbourhood formation

process, the user profiles for the reference users are required to be stored in the cache.

The memory requirement for the reference user profiles is trivial, because there are only

three reference users.

Page 216

Figure C.4. An example structure of the RDF searching cache

Given that the RDF searching cache is properly initialised, the detailed RDF

procedure is described below:

Algorithm C.1 , ,

Input is the target user whose neighbourhood is to be formed.

is the overall user set which is the target search space.

is the target neighbourhood size.

Output Neighbour is ’s neighbourhood in

1) With the proposed RDF searching cache, use the indexed tree structure to locate the

minimal user nodes set within the given boundary:

Page 217

| , | , |

where , , is one of the reference users, and is set such that the

estimated searching space is minimal. Also, it can be observed that the equation

for computing depicted here is based on Equation (C.1), and we rewrote this

equation here to accommodate the new notations used in this section for describing

the RDF searching cache.

2) Based on step 1, , , is the primary index key used for sorting and

retrieving ξ . The rest two index keys (i.e. , , \ ) are denoted as and

.

3) FOR EACH

4) IF , or | , |

THEN remove from

5) END FOR

6) Do the standard ‘best- -neighbours’ search against the estimated searching space ,

and return the result neighbourhood for (the final neighbourhood size is limited

to smaller than or equal to ).

It can be seen from line (3) to line (5) of the Algorithm C.1, the size of the

searching space is further reduced by using reference users and . This process is

similar to finding the intersected space as described in Section C.1.

C.3. EXPERIMENTS AND EVALUATION

The goal of the experiment presented in this section is to evaluate whether the

proposed RDF technique can effectively improve the recommendation performance and

Page 218

computation efficiency of recommenders. Hence, this experiment involves a standard

recommender system and a set of baseline neighbourhood formation techniques. By

observing the recommender’s performance affected by equipping it with different

neighbourhood formation techniques, we can evaluate whether the proposed RDF

technique is indeed effective.

The recommender system employed in this experiment is the Taxonomy Product

Recommender (TPR) proposed by Ziegler et al. (2004), for detailed information about

this technique please refer to Section 2.2, Section 3.3.3.1 and (Ziegler et al., 2004).

C.3.1. Data Acquisition

The dataset employed in this experiment is the ‘Book-Crossing’ dataset

(http://www.informatik.uni-freiburg.de/~cziegler/BX/) which is also the main

experiment dataset employed in Chapter 3. Please refer to Section 3.3.1 for more details

about the dataset.

Because the TPR uses only implicit user ratings, therefore, we further removed

all explicit user ratings from the dataset and kept the remaining 716,109 implicit ratings

for the experiment.

C.3.2. Experiment Framework

In order to evaluate whether the proposed RDF method is effective in improving

recommenders’ recommendation quality and computation efficiency, we implemented

four different version of TPR, and each of them is equipped with different

neighbourhood formation algorithms. The four TPR versions are:

Page 219

TPR: this is the standard TPR version, and there is no optimisation in its

neighbourhood formation process. That is, the neighbourhood formation

process involved requires comparing the target user with all other users in

the dataset.

RDF based TPR: this TPR version employs the proposed RDF method to

form the neighbourhood.

RTree based TPR: this TPR version employs the RTree (Manolopoulos et

al., 2005) technique to form the neighbourhood. RTree is a tree structure

based neighbourhood formation method, and it has been widely applied in

many applications.

Random TPR: this TPR version forms its neighbourhood with randomly

chosen users. It is used as the baseline for the recommendation quality

evaluation.

C.3.3. Evaluation Metrics

In the recommendation performance part of evaluation, the k-folding technique

(Herlocker et al., 2004) is employed (where is set to 5 in our setting). With k-folding,

every user  ’s implicit rating list is divided into 5 equal size portions. With these

portions, one of them is selected as  ’s training set , and the rest 4 portions are

combined into a test set \ . Totally we have five combinations , ,

1 5 for user   . In the experiment, the recommenders will use the training set

to learn  ’s interest, and the recommendation list generated for   will then be

evaluated according to . Moreover, the size for the neighbourhood formation is set to

20 and the number of items within each recommendation list is set to 20 too.

Page 220

The metrics used in this experiment to evaluate the recommendation list (i.e. )

against the testing item list (i.e. ) are the Precision metrics and Recall metrics, for

detailed information about these two metrics please refer to Equation (3.19) and

Equation (3.20).

For the computation efficiency evaluation, the average time required by different

TPRs to make a recommendation will be compared. We incrementally increase the

number of users in the dataset (from 1000, 2000, 3000 until 14000), and observe how the

computation times are affected by the increments.

C.3.4. Experimental Results

Figure C.5 and Figure C.6 show the performance comparison between the

standard TPR and the proposed RDF based TPR using the precision and recall metrics.

The horizontal axis for both precision and recall charts indicates the minimum number of

ratings in the user’s profile (i.e.| |). Therefore, larger x-coordinates imply that fewer

users are considered for the evaluation. It can be observed that the proposed RDF based

TPR outperformed standard TPR for both recall and precision. The result confirms that

when the dissimilar users are removed from the neighbourhood, the quality of the result

recommendations become better. RTree based TPR performs much worse than both the

RDF based TPR and the standard TPR, as it is unable to accurately allocate neighbours

for target users.

Page 221

Figure C.5. Precision Results for different TPR versions

Figure C.6. Recall Results for different TPR versions

The efficiency evaluation is shown in Figure C.7. It can be seen from Figure C.7

that the time efficiency for standard TPR drops drastically when the number of users in

the dataset increases. For dataset with 15000 users, the system needs about 14 seconds to

produce a recommendation for a user, and it is not acceptable for most commercial

systems. By comparison, the RDF based TPR is much efficient, and it only needs less

Page 222

than 4 seconds to produce a recommendation for dataset with 15000 users. The RTree

based TPR greatly outperforms the proposed method when the number of users in the

dataset is under 8000. However, as the number of users increases in the dataset, the

differences between RDF and RTree based TPR becomes smaller, and RDF starts

outperforms RTree when the number of users in the dataset is over 9000. This is because

RTree is only efficient when the tree level is small. However, as the tree level increases

(i.e. when number of users increases) RTree’s performance drops drastically because the

chance for high dimensional vector comparison increases quadratically in accordance to

the number of tree level. The proposed RDF method outperforms RTree method because

its indexing strategy is single value based, and it reduces the possibility for the high

dimensional vector correlation computation.

Figure C.7. Average recommendation time for different TPR versions

Page 223

Bibliography

ADOMAVICIUS, G., SANKARANARAYANAN, R., SEN, S. & TUZHILIN, A.

(2005) Incorporating contextual information in recommender systems using a

multidimensional approach ACM Trans. Inf. Syst., 23, 103-145.

AL-DAOUD, M. D. B. (2005) A New Algorithm for Cluster Initialization. Transactions

on Engineering Computing and Technology. Istanbul, Turkey.

AROYO, L., STASH, N., WANG, Y., GORGELS, P. & RUTLEDGE, A. L. (2007)

CHIP demonstrator: semantics-driven recommendations and museum tour

generation. Semantic Web Challenge 2007. Busan, Korea.

AWERBUCH, B., PATT-SHAMIR, B., PELEG, D. & TUTTLE, M. (2005) Improved

recommendation systems. 16th annual ACM-SIAM Symposium on Discrete

algorithms. Vancouver, British Columbia.

AZOULAY-SCHWARTZ, R., KRAUS, S. & WILKENFELD, J. (2004) Exploitation vs.

exploration: choosing a supplier in an environment of incomplete information.

Decision Support Systems, 38, 1--18.

BADRUL, S., GEORGE, K., JOSEPH, K. & JOHN, R. (2001) Item-based collaborative

filtering recommendation algorithms. Proceedings of the 10th international

conference on World Wide Web. Hong Kong, Hong Kong, ACM.

BALABANOVIĆ, M. & SHOHAM, Y. (1997) Fab: content-based, collaborative

recommendation Communications of the ACM, 40, 66-72.

BASU, C., HIRSH, H. & COHEN, W. W. (1998) Recommendation as classification:

Using social and content-based information in recommendation. 5th National

Conference on Artificial Intelligence.

Page 224

BERKHIN, P. (2002) Survey Of Clustering Data Mining Techniques. San Jose, CA,

Accrue Software.

BILLSUS, D., PAZZANI, M. J. & CHEN, J. (2000) A learning agent for wireless news

access. 5th international conference on Intelligent user interfaces New Orleans,

Louisiana, United States

BILLSUS, D. & PAZZANI., M. (1999) A hybrid user model for news classification. 7th

International Conference on User Modelling. New York, Spring-Verlag.

BOONE, G. (1998) Concept features in Re:Agent, an intelligent Email agent 2nd

international conference on Autonomous agents Minneapolis, Minnesota, United

States

BREESE, J. S., HECKERMAN, D. & KADIE, C. (1998) Empirical Analysis of

Predictive Algorithms for Collaborative Filtering. Proceedings of 14th

Conference on Uncertainty in Artificial Intelligence. Madison, WI.

BURKE, R. (2002) Hybrid Recommender Systems: Survey and Experiments. User

Modeling and User-Adapted Interaction, 12, 331-370.

CASTAGNOS, S. & BOYER, A. (2007) Modeling Preferences in a Distributed

Recommender System. Lecture Notes in Computer Science. Springer Berlin /

Heidelberg.

CHEN, J. R., WOLFE, S. R. & WRAGG, S. D. (2000) A distributed multi-agent system

for collaborative information management and sharing. 9th International

Conference on Information and Knowledge Management. McLean, Virginia,

United States, ACM.

CHRISTOPH, B. (1997) A probabilistic model for distributed information retrieval. 20th

annual international ACM SIGIR conference on Research and development in

information retrieval. Philadelphia, Pennsylvania, United States, ACM.

Page 225

CLEMENTS, M., VRIES, A. P. D., POUWELSE, J. A., WANG, J. & REINDERS, M. J.

T. (2007) Evaluation of Neighbourhood Selection Methods in Decentralized

Recommendation Systems. Workshop on Large Scale Distributed Systems for

Information Retrieval Netherlands.

CLEVERDON, C. W., MILLS, J. & KEEN, M. (1966) Factors determining the

performance of indexing systems. ASLIB Cranfield project, Cranfield.

COHEN, W. W. (1995) Fast Effective Rule Induction. IN PRIEDITIS, A. & RUSSELL,

S. (Eds.) 12th International Conference on Machine Learning. Tahoe City, CA,

Morgan Kaufmann.

COHEN, W. W. (1996) Learning rules that classify e-mail. AAAI Spring Symposium on

Machine Learning in Information Access.

COOLEY, R., TAN, P.-N. & SRIVASTAVA, J. (1999) Websift: the web site

information filter system. 1999 KDD Workshop on Web Mining. San Diego, CA,

Springer-Verlag.

CÖSTER, R., GUSTAVSSON, A., OLSSON, T. & RUDSTRÖM, Å. (2002) Enhancing

web-based configuration with recommendations and cluster-based help.

Workshop on Recommendation and Personalization in eCommerce. Malaga,

Spain.

CUNNINGHAM, P., BERGMANN, R., SCHMITT, S., TRAPHÖNER, R., BREEN, S.

& SMYTH, B. (2001) WEBSELL: Intelligent Sales Assistants for the World

Wide Web. KI - Zeitschrift fr Knstliche Intelligenz.

DEGEMMIS, M., LOPS, P., SEMERARO, G., COSTABILE, M. F., GUIDA, S. P. &

LICCHELLI, O. (2004) Improving Collaborative Recommender Systems by

means of User Profiles. Human-Computer Interaction Series: Designing

personalized user experiences in eCommerce. Norwell, MA, USA, Kluwer

Academic Publishers.

Page 226

DESHPANDE, M. & KARYPIS, G. (2004) Item-based top-N recommendation

algorithms. ACM Transactions on Information Systems, 22, 143-177.

DRINEAS, P., KERENIDIS, I. & RAGHAVAN, P. (2002) Competitive

recommendation systems. 34th annual ACM symposium on Theory of computing.

New York, NY, USA, ACM Press.

FERMAN, A. M., ERRICO, J. H., BEEK, P. V. & SEZAN, M. I. (2002) Content-based

filtering and personalization using structured metadata. 2nd ACM/IEEE-CS joint

conference on Digital libraries Portland, Oregon, USA.

FONER, L. N. (1997) Yenta: a multi-agent, referral-based matchmaking system. 1st

International Conference on Autonomous agents. Marina del Rey, California,

United States, ACM.

FRENCH, J. C., POWELL, A. L., CALLAN, J. P., VILES, C. L., EMMITT, T., PREY,

K. J. & MOU, Y. (1999) Comparing the Performance of Database Selection

Algorithms. Research and Development in Information Retrieval.

FRIGUI, H. & KRISHNAPURAM, R. (1997) Clustering by competitive agglomeration.

Pattern recognition, 30, 1109-1119

FUNAKOSHI, K. & OHGURO, T. (2000) A content-based collaborative recommender

system with detailed useof evaluations. 4th Conference on Knowledge-Based

Intelligent Engineering Systems and Allied Technologies.

GHANI, R. & FANO, A. (2002) Building recommender systems using a knowledge

base of product semantics. Workshop on Recommendation and Personalization

in E-Commerce (RPEC). Malaga, Spain, Springer-Verlag.

GOLDBERG, D., NICHOLS, D., OKI, B. M. & TERRY, D. (1992) Using collaborative

filtering to weave an information tapestry. Communications of the ACM, 35, 61-

70.

Page 227

GOOD, N., SCHAFER, J. B., KONSTAN, J. A., BORCHERS, A., SARWAR, B. M.,

HERLOCKER, J. L. & RIEDL, J. (1999) Combining collaborative filtering with

personal agents for better recommendations. 6th National Conference on

Artificial Intelligence.

GUI-RONG, X., CHENXI, L., QIANG, Y., WENSI, X., HUA-JUN, Z., YONG, Y. &

ZHENG, C. (2005) Scalable collaborative filtering using cluster-based

smoothing. 28th annual international ACM SIGIR conference on research and

development in information retrieval. Salvador, Brazil, ACM.

HAN, P., XIE, B., YANG, F. & SHEN, R. (2004) A scalable P2P recommender system

based on distributed collaborative filtering. Expert Systems with Applications, 27,

203-210.

HAYES, C., MASSA, P., AVESANI, P. & CUNNINGHAM, P. (2002) An on-line

evaluation framework for recommender systems. Personalization and

Recommendation in E-Commerce. Malaga.

HERLOCKER, J., KONSTAN, J. A. & RIEDL, J. (2002) An empirical analysis of

design choices in neighborhood-based collaborative filtering algorithms.

Information Retrieval, 5, 287-310.

HERLOCKER, J. L., KONSTAN, J. A., TERVEEN, L. G. & RIEDL, J. T. (2004)

Evaluating collaborative filtering recommender systems. ACM Transactions on

Information Systems (TOIS), 22, 5-53.

HOLLINK, L., SCHREIBER, G. & WIELINGA., B. (2007) Patterns of semantic

relations to improve image content search. Journal of Web Semantics, 5, 195-203.

JAIN, A. K., MURTY, M. N. & FLYNN, P. J. (1999) Data Clustering: A Review. ACM

Computing Surveys, 31, 264-323.

JENNINGS, A. & HIGUCHI, H. (1993) A user model neural network for a personal

news service User Modeling and User-Adapted Interaction, 3, 1-25.

Page 228

JEROME, K. & DEREK, B. (2004) An accurate and scalable collaborative

recommender. Artif. Intell. Rev., 21, 193-213.

JIAN, C., JIAN, Y. & JIN, H. (2005) Automatic content-based recommendation in e-

commerce. e-Technology, e-Commerce and e-Service.

JOHN, G. (1989) Multi-armed bandit allocation indices, Wiley.

JUN, W., ARJEN, P. D. V. & MARCEL, J. T. R. (2006) Unifying user-based and item-

based collaborative filtering approaches by similarity fusion. Proceedings of the

29th annual international ACM SIGIR conference on Research and development

in information retrieval. Seattle, Washington, USA, ACM.

KARYPIS, G. (2001) Evaluation of Item-Based Top-N Recommendation Algorithms.

10th Conference of Information and Knowledge Management.

KIM, J. W., LEE, B. H., SHAW, M. J., CHANG, H.-L. & NELSON, M. (2001)

Application of decision-tree induction techniques to personalized advertisements

on internet storefronts. International Journal of Electronic Commerce 5, 45-62.

KOHRS, A. & MERIALDO, B. (2000) Using category-based collaborative filtering in

the active WebMuseum. IEEE International Conference on Multimedia and

Expo.

KONSTAN, J. A., MILLER, B. N., MALTZ, D., HERLOCKER, J. L., GORDON, L. R.

& RIEDL, J. (1997) GroupLens: applying collaborative filtering to Usenet news.

Communications of the ACM, 40, 77-87.

KRETSER, O. D., MOFFAT, A., SHIMMIN, T. & ZOBEL, J. (1998) Methodologies

for Distributed Information Retrieval. International Conference on Distributed

Computing Systems.

KRISS, S. (2007) Collaborative Filtering and the Netflix Challenge. Yale University.

Page 229

KRULWICH, B. (1997) LIFESTYLE FINDER: Intelligent User Profiling Using Large-

Scale Demographic Data. AI Magazine, 18, 37-45.

LEMIRE, D. & MACLACHLAN, A. (2005) Slope One Predictors for Online Rating-

Based Collaborative Filtering. 2005 SIAM Data Mining

LEO, O., HOWARD, H. L. & ROBERT, E. W. (2003) Ontologies for corporate web

applications. AI Mag., 24, 49-62.

LEVY, T. (2004) The state and value of taxonomy standards. The Seybold Report July

21, 2004.

LIKAS, A., VLASSIS, N. & VERBEEK, J. J. (2003) The Global K-means Clustering

Algorithm. Pattern Recognition, 36, 451-461.

LINDEN, G., SMITH, B. & YORK, J. (2003) Amazon.com recommendations: item-to-

item collaborative filtering. Internet Computing, IEEE, 7, 76-80.

LINK, H., SAIA, J., LANE, T. & LAVIOLETTE, R. A. (2005) The Impact of Social

Networks on Multi-Agent Recommender Systems. CoRR, abs/cs/0511011.

LIU, P., NIE, G., CHEN, D. & FU, Z. (2007) The Knowledge Grid Based Intelligent

Electronic Commerce Recommender Systems. IEEE International Conference

on Service-Oriented Computing and Applications. Newport Beach, CA, USA.

MALONE, T. W., GRANT, K., TURBAK, F., A.BROBST, S. & COHEN, M. D. (1987)

Intelligent information-sharing systems. Communications of the ACM, 30, 390-

402.

MANOLOPOULOS, Y., NANOPOULOS, A., PAPADOPOULOS, A. N. &

THEODORIDIS, Y. (2005) R-Trees: Theory and Applications, Springer.

MIDDLETON, S. E., ALANI, H., SHADBOLT, N. R. & DE ROURE, D. C. (2002)

Exploiting Synergy Between Ontologies and Recommender Systems. The

Semantic Web Workshop, World Wide Web Conference.

Page 230

MIDDLETON, S. E., SHADBOLT, N. R. & ROURE, D. C. D. (2004) Ontological User

Profiling in Recommender Systems. ACM Transactions on Information Systems,

22, 54-88.

MILLER, B. N., KONSTAN, J. A. & RIEDL, J. (2004) PocketLens: Toward a personal

recommender system. ACM Transactions on Information Systems, 22, 437-476.

MIN, S.-H. & HAN, I. (2005) Recommender systems using support vector machines.

International Conference on Web Engineering.

MIRA, K. & DONG-SUB, C. (2001) Collaborative filtering with automatic rating for

recommendation. IN DONG-SUB, C. (Ed. IEEE International Symposium on

Industrial Electronics.

MLADENIC, D. (1996) Personal WebWatcher: design and implementation. Technical

Report IJS-DP-7472. Pittsburgh, USA, School of Computer Science, Carnegie-

Mellon University.

MONTANER, M., LÓPEZ, B. & ROSA, J. L. D. L. (2003) A Taxonomy of

Recommender Agents on the Internet. Artificial Intelligence Review, 19, 285-330.

OGSTON, E., OVEREINDER, B., STEEN, M. V. & BRAZIER, F. (2003) A method

for decentralized clustering in large multi-agent systems. 2nd International Joint

Conference on Autonomous Agents and Multiagent Systems. Melbourne,

Australia, ACM.

PAPAGELIS, M. & PLEXOUSAKIS, D. (2004) Qualitative Analysis of User-Based

and Item-Based Prediction Algorithms for Recommendation. Lecture Notes in

Computer Science, 3191/2004, 152-166.

PAPAGELIS, M., ROUSIDIS, I. & PLEXOUSAKIS, D. (2005) Incremental

Collaborative Filtering for Highly-Scalable Recommendation Algorithms.

Proceedings of the15th International Symposium on Methodologies of Intelligent

Systems.

Page 231

PARK, S.-T., PENNOCK, D., MADANI, O., GOOD, N. & DECOSTE, D. (2006)

Naive filterbots for robust cold-start recommendations. IN PRESS, A. (Ed. 12th

ACM SIGKDD international conference on Knowledge discovery and data

mining. Philadelphia, PA, USA.

PAZZANI, M., MURAMATSU, J. & BILLSUS, D. (1996) Syskill & Webert:

Identifying interesting web sites. 13th National Conference on Artificial

Intelligence.

PAZZANI, M. J. (1999) A framework for collaborative, content-based and demographic

filtering Artificial Intelligence Review, 13, 393-408.

PAZZANI, M. J. & BILLSUS, D. (2007) Content-based recommender systems. IN

BRUSILOVSKY, P., KOBSA, A. & NEJDL, W. (Eds.) The Adaptive Web.

Berlin, Germany, Springer-Verlag.

PEDRYCZ, W. (2005) Clustering and Fuzzy Clustering. Knowledge-Based Clustering:

From Data to Information Granules. Wiley InterScience.

PELLEG, D. & MOORE, A. (1999) Accelerating Exact k-means Algorithms with

Geometric Reasoning. Knowledge Discovery and Data Mining.

PELLEG, D. & MOORE, A. (2000) X-means: Extending K-means with Efficient

Estimation of the Number of Clusters. Seventeenth International Conference on

Machine Learning. San Francisco, Morgan Kaufmann.

POPESCUL, A., UNGAR, L., PENNOCK, D. & LAWRENCE, S. (2001) Probabilistic

Models for Unified Collaborative and Content-Based Recommendation in

Sparse-Data Environments. 17th Conference on Uncertainty in Artificial

Intelligence.

PRETSCHNER, A. & GAUCH, S. (1999) Ontology based personalized search. 11th

IEEE International Conference on Tools with Artificial Intelligence. IEEE

Computer Society.

Page 232

RASHID, A. M., LAM, S. K., LAPITZ, A., KARYPIS, G. & RIEDL, J. (2006a)

ClustKNN: a highly scalable hybrid model- and memory-based CF algorithm.

Workshop on Web Mining and Web Usage Analysis. Philadelphia, Pennsylvania.

RASHID, A. M., LAM, S. K., LAPITZ, A., KARYPIS, G. & RIEDL, J. (2006b)

Towards a Scalable kNN CF Algorithm: Exploring Effective Applications of

clustering. Workshop on Web Mining and Web Usage Analysis. Philadelphia,

Pennsylvania.

RESNICK, P. & VARIAN, H. R. (1997) Recommender systems. Communications of

the ACM 40, 56--58.

RICH, E. (1998) User modeling via stereotypes. Readings in intelligent user interfaces.

San Francisco, CA, USA, Morgan Kaufmann Publishers Inc.

RUSSELL, S. & NORVIG, P. (2002) Artificial Intelligence: A Modern Approach,

Prentice.

SAITOH, S. (2003) Generalizations of the Triangle Inequalty. Journal of Inequalities in

Pure and Applied Mathematics, 4.

SALTON, G. (1983) Introduction to Modern Information Retrieval, New York,

McGraw-Hill Companies.

SARWAR, B., KARYPIS, G., KONSTAN, J. & RIEDL, J. (2000a) Application of

dimensionality reduction in recommender systems--a case study. ACM WebKDD

Workshop. Boston, MA, USA.

SARWAR, B., KARYPIS, G., KONSTAN, J. & RIEDL, J. (2002) Recommender

systems for large-scale e-commerce: Scalable neighborhood formation using

clustering. Fifth International Conference on Computer and Information

Technology.

Page 233

SARWAR, B. M., KARYPIS, G., KONSTAN, J. A. & RIEDL, J. (2000b) Analysis of

recommendation algorithms for e-commerce. ACM Conference on Electronic

Commerce.

SCHAFER, J. B., KONSTAN, J. A. & RIEDL, J. (2000) E-Commerce

Recommendation Applications. Journal of Data Mining and Knowledge

Discovery, 5, 115-152.

SCHEIN, A. I., POPESCUL, A., UNGAR, L. H. & PENNOCK, D. M. (2002) Methods

and metrics for cold-start recommendations 25th annual international ACM

SIGIR conference on Research and development in information retrieval.

Tampere, Finland, ACM Press.

SCHWAB, I., POHL, W. & KOYCHEV, I. (2000) Learning to recommend from

positive evidence. AAAI 2000 Spring Symposium: Adaptive User Interface.

SHARDANAND, U. & MAES, P. (1995) Social information filtering: algorithms for

automating word of mouth. CHI'95 Conference on Human Factors in

Computing Systems. ACM Press.

SMITH, R. G. (1981) The Contract Net Protocol: High-Level Communication and

Control in a Distributed Problem Solver. IEEE Transactions on Computers, C-

29, 1104--1113.

SOLLENBORN, M. & FUNK, P. (2002) Category-based filtering and user stereotype

cases to reduce the latency problem in recommender systems. 6th European

Conference on Advances in Case-Based Reasoning London, UK, Springer-

Verlag.

SORGE, C. (2007) A Chord-based Recommender System. 32nd IEEE Conference on

Local Computer Networks, 2007. LCN 2007. .

TABACHNICK, B. G. & FIDELL, L. S. (2006) Using Multivariate Statistics, Allyn &

Bacon.

Page 234

TERVEEN, L., HILL, W., AMENTO, B., MCDONALD, D. & CRETER, J. (1997)

PHOAKS: A system for sharing recommendations. Communications of the ACM,

40, 59-62.

TOWLE, B. & QUINN, C. (2000) Knowledge based recommender systems using

explicit user models. Knowledge-Based Electronic Markets Workshop at AAAI

2000. Austin, TX.

TVEIT, A. (2007) Peer-to-peer based Recommendations for Mobile Commerce. the

First International Workshop on Mobile Commerce. Rome, Italy.

VIDAL, J. M. V. M. (2004) A Protocol for a Distributed Recommender System.

Trusting Agents for Trusting Electronic Societies.

WANG, J., POUWELSE, J., LAGENDIJK, R. & REINDERS, M. R. J. (2006)

Distributed Collaborative Filtering for Peer-to-Peer File Sharing Systems. 21st

Annual ACM Symposium on Applied Computing.

WEI, Y. Z., MOREAU, L. & JENNINGS, N. R. (2003) Recommender Systems: A

Market-Based Design. 2nd International Joint Conference on Autonomous

Agents and Multiagent Systems. Melbourne, Australia.

WEI, Y. Z., MOREAU, L. & JENNINGS, N. R. (2005) A market-based approach to

recommender systems. ACM Transactions on Information Systems 23, 227-266.

WEISS, G. (1999) Multiagent Systems: a modern appraoch to distributed artificial

intelligence, London, England, The MIT Press.

XU, Y. (2005) Hybrid Clustering with Application to Web Mining. Active Media

Technology. Japen.

YANG, C. C., CHEN, H. & HONG, K. (2003) Visualization of large category map for

Internet browsing Decision Support Systems, 35, 89-102.

Page 235

YANG, J., WANG, J., CLEMENTS, M., POUWELSE, J. A., VRIES, A. P. D. &

REINDERS, M. (2007) An Epidemic-based P2P Recommender System.

Workshop on Large Scale Distributed Systems for Information Retrieval

Netherlands.

ZENG, C., XING, C.-X. & ZHOU, L.-Z. (2003) Similarity measure and instance

selection for collaborative filtering. Proceedings of the 12th international

conference on World Wide Web. Budapest, Hungary

ZIEGLER, C.-N. & GOLBECK, J. (2007) Investigating interactions of trust and interest

similarity. Decision Support Systems, 43, 460-475.

ZIEGLER, C.-N., LAUSEN, G. & SCHMIDT-THIEME, L. (2004) Taxonomy-driven

Computation of Product Recommendations International Conference on

Information and Knowledge Management Washington D.C., USA