Using an Ontological A-priori Score to Infer User’s Preferences

Vincent Schickel-Zuber - AI Lab 29th August 2006

Using an Ontological A-priori Score to Infer User’s Preferences

W17: Workshop on Recommender Systems – ECAI 2006

Advisor: Prof Boi Faltings – EPFL

Vincent Schickel-Zuber - AI Lab 2

Presentation Layout1. Introduction

Introduce the problem and existing techniques2. Transferring User’s Preference

Introduce the assumptions behind our model Explain the transfer of preference

3. Validation of the model Experiment on MovieLens

4. Conclusion Remarks & Future work


Problem Definition Recommendation Problem (RP):

Recommend a set of items I to the user from a set of all items O, based on his preferences P. Use a Recommender System, RS, to find the best

items Examples:

NotebookReview.com (O=Notebooks, P= criteria (Processor Type, Screen Size))

Amazon.com (O=Books, DVDs,… , P= grading) Google (O=Web Documents, P= keywords)


Recommendation Systems Three approaches to build a RS: [1][2][3][4][5]

1. Case-Based Filtering: uses previous casesi.e.: Collaborative Filtering (cases – user’s ratings)

Good performances – low cognitive requirementsSparsity, latency, shilling attacks and cold start problem

3. Rule-Based Filtering: uses association between itemsi.e.: Data Mining (associations – rules)

Find hidden relationships – good domain discoveryExpensive and time consuming

2. Content-Based Filtering: uses item’s descriptioni.e.: Multi-Attribute Utility Theory (descriptions-attributes)

Match user’s preferences – very good precisionElicitation of weights and value function.


Central Problem of RS

A Major Problem in RS: The Elicitation Problem=> Incomplete user’s model

Collaborative Filtering

5I44I55

3I245

4I134

Multi-Attribute Utility Theory




Introduce the assumption behind our model Explain the transfer of preference

3. Validation of the model Experiment on MovieLens



OntologyD1 Ontology λ is a graph (DAG) where

nodes models concepts Instances being the items

edges represents the relations (features). Sub-concepts are distinguished by certain features Feature are usually not made explicit

Car

Vehicle

Transport

Boat

Bus

On-land On-sea

<7 >6

Compact SUV

City All_terrain


The Score of Concept -S The RP viewed as predicting the score S assigned

to a concept (group of items).

The score can be seen as a lower bound function that models how much a user likes an item

S is a function that satisfies the assumptions: A1: S depends on the features of the item

Items are models by a set of features A2: Each feature contributes independently to S

Eliminates the inter-dependence between features A3: unknown|disliked features make no contribution

Reflects the fact that users are risk-averse Liking a concept liking a sub-concept


A-priori Score - APS The structure of the ontology contains information

Use APS(c) to capture the knowledge of concept c If no information, assume S(c) uniform [0..1]

P(S(c)>x)=1-x Concepts can have n descendants

Assumption A3 => P(S(c)>x)=(1-x)n+1

E(c)= ∫xfc(x)dx = 1n+2

APS(c)= n+21

#descendantsAP

S

0,5

leafs

root

APS uses no user information


Inference Idea

Car

Vehicle

Bus

S(SUV)=0.8

SUV

S(bus)=???

Select the best Lowest Common Ancestor lca(SUV, bus) – AAAI’06


Upward Inference

Going up k levels ⇒ remove k known features

A1 the score depends on the features of the item K levels

SUV

vehicle

Removing features ⇒ S↘ or S ↔ (S =∑S) S( vehicle | SUV)= α( vehicle, SUV) * S(SUV)

α ∈[0..1] is the ratio of feature in common liked

How to compute α?α =#feature(vehicle) / #feature(SUV) Does not take into account the feature distribution

α =APS(vehicle) / APS(SUV)


Downward Inference

Going down l levels ⇒ adding l unknown features

l levelsbus

vehicle

Adding features ⇒ S↗ or S↔ (S =∑S)S(bus|vehicle)=α S(vehicle) α ≥ 1

How to compute β? β = APS(bus) - APS(vehicle)

⇏ S(bus|vehicle)= S(vehicle) + β(vehicle, bus)

β ∈[0..1] is ∑features in bus not present in vehicle

A3 Users are pessimistic liking some features liking othersA2 Features contributes independently to the score


Overall Inference

Car

Vehicle

SUV

Bus

There exist a chain between “city” and vehicle but not a path

As for Bayesian Networks, we assume independence

S(Bus|SUV)= αS(SUV) + β

The score of a concept x knowing y is defined as:S(y|x)= α(x,lcax,y)S(x) + β(y,lcax,y)

Use APS The score function is asymmetric

Elicited from the user




Introduce the assumption behind our model Explain the transfer of preference

3. Validation of the model WordNet (built best similarity metric – see paper) Experiment on MovieLens



Validation – Transfer - I

MovieLens – movies are modeled by 23 Attributes 19 themes, MPPA rating, duration, and released date. Extracted from IMDB.com

MovieLens database used by CF community: 100,000 ratings on 1682 movies done by 943 users.

Built an ontology modeling the 22 attributes of a movies Used definitions found in various online dictionaries

-name

Themes

Unknown

Animation Adventure DocumentaryDrama

Musical

Comedy

Romance

Scifi

Action

Mystery

Western War

Horror

Thriller

Crime

Film-Noir

Children Fantasy

null

1 *


Validation – Transfer - II Experiment Setup – for each 943 users

1. Filtered users with less than 65 ratings2. Split user’s data into learning set and test set 3. Computed utility functions from learning set

1. Frequency count algorithm for only 10 attributes2. Our inference approach for other 12 attributes

4. Predicted the grade of 15 movies from the test set Our approach – HAPPL (LNAI 4198 – WebKDD’05) Item-Item based CF (using adjusted Cosine) Popularity ranking

5. Computed the accuracy of predictions for Top 5 Used the Mean Absolute Error (MAE)

6. Back to 3 with a bigger training set {5,10,20,…,50}


Validation – Transfer - IIITop 5 Strategy

0.73

0.78

0.83

0.88

0.93

0.98

5 10 20 30 40 50

#learning ratings in learning set

MA

E

Popularity

Hybrid

HAPPL

CF


Validation – Transfer - IVTop 5 Strategy

0

0.05

0.1

0.15

0.2

0.25

0.3

0.35

0.4

5 10 20 30 40 50

#learning ratings in learning set LS

Nov

elty

HAPPL| Popularity

Hybrid | Popularity

CF| Popularity


Conclusions

Requirements & Conditions: A2 - Features contributes to preference independent. Need an ontology modeling all the domain

Next steps: Try to learn the ontology Preliminary results shows that we still outperform CF Learn ontology gives a more restricted search space

We have introduced the idea that ontology could be used to transfer missing preferences. Ontology can be used to compute A-priori score Inference model - asymmetric property Outperforms CF without other people information

n+21


Questions?

Thank-youSlides: http://people.epfl.ch/vincent.schickel-zuber

http://people.epfl.ch/vincent.schickel-zuber


References - I[1] Survey of Solving Multi-Attribute Decisions Problems

Jiyong Zang, and Pearl Pu, EPFL Technical Report, 2004.

[2] Improving Case-Based Recommendation A Collaborative Filtering ApproachDerry O’Sullivan, David Wilson, and Barry Smyth, Lecture Notes In Computer Science, 2002.

[3] An improved collaborative Filtering approach for predicting cross-category purchases based on binary market data.Andreas Mild, and Thomas Reutterer, Journal of Retailing and Consumer Services Special Issue on Model Building in Retailing & consumer Service, 2002.

[4] Using Content-Based Filtering for RecommendationRobin van Meteren and Maarten van Someren, ECML2000 Workshop, 2000.

[5] Content-Based Filetering and Personalization Using Structure MetadataA. Mufit Ferman, James H. Errico, Peter van Beek, and M Ibrahim Sezan, JCDL02, 2002.


References - II[AAAI’06] Inferring User’s Preferences Using Onotlogies

Vincent Schickel and Boi Faltings, In Proc. AAAI’06 pp 1413 – 1419, 2006.

[LNAI 4198] Overcoming Incomplete User Models In Recommendation Systems via an Ontology.Vincent Schickel and Boi Faltings, LNAI 4198, pp 39 -57, 2006.