1 Web Usage Mining Modelling: frequent-pattern mining I (sequence mining with WUM), classification and clustering) Prof. Dr. Bettina Berendt Humboldt Univ.

1

Web Usage Mining

Modelling: frequent-pattern mining I (sequence mining with WUM), classification and clustering)

Prof. Dr. Bettina BerendtHumboldt Univ. Berlin, Germany

www.berendt.de

2

Please note

These slides use and/or refer to a lot of material available on the Internet. To reduce clutter, credits and hyperlinks are given in the following ways:

Slides adapted from other people‘s materials: at bottom of slide

Pictures, screenshots etc.: URL visible in screenshot or given in PPT „Comments“ field

Literature, software: On the accompanying Web site http://vasarely.wiwi.hu-berlin.de/WebMining07/

Thanks to the Internet community!

You are invited to re-use these materials, but please give the proper credit.

3

Stages of knowledge discoverydiscussed in this lecture

Applicationunderstanding

4An addendum to the association rules:main interestingness measures of association rules(and a recommendation for postprocessing the result set)

Support of a rule A B

= no. of instances with A and B / no. of all instances

Confidence of a rule A B

= no. of instances with A and B / no. of instances with A

= support (A & B) / support (A)

Lift of a rule A B

= support (A & B) / [ support (A) * support (B) ]

What does this measure, and in what numerical interval can it be?

Deleting redundant rules from the result set:

If you have A B and A & C B, the second rule is redundant.

5Agenda

Sequence mining: tool WUM (case study “school search”)

Classification: method Naïve Bayes (case study “happiness”)

Clustering: tool DocumentAtlas (case study “EU proposals”)

A very short note on other uses of clustering (e.g. in query mining)

Some observations on privacy ...

Best-practice „design patterns“with open-source tools

6

Demonstration of WUM

7

The site

Business understanding / problem definition:

* How do users search in this online catalog?

* Which search criteria are popular?

* Which are efficient?

[Berendt & Spiliopoulou,VLDB Journal 2000]

8

The concept hierarchies / site ontology(excerpt)

SEITE1-...LI (1st page of a list)orSEITEn-...LI (further page)

LA („Land“) SA („Schulart“) SU („Suche“)

9Sequence mining – one result pattern: successful search for a school in Germany

a refinement

a repetition

a continuation

one example pattern

select t from node a b, template a * b as t where a.url startswith "SEITE1-" and a.occurrence = 1 and b.url contains "1SCHULE" and b.occurrence = 1 and (b.support / a.support) >= 0.2

(Berendt & Spiliopoulou, VLDB J. 2000)

/liste.html?offset=920&zeilen=20&anzahl=1323&sprache=de&sw_kategorie=de&erscheint=&suchfeld=&suchwert=&staat=de&region=by&schultyp=

/liste.html?offset=920&zeilen=20&anzahl=1323&sprache=de&sw_kategorie=de&erscheint=&suchfeld=&suchwert=&staat=de&region=by&schultyp=

10

Sequences

11Generalized sequences, navigation patterns, hits in WUM

12Aggregated Logs: The basic internal representation in WUM

13The confi-dence measure for genera-lized sequences

14Templates in the query language MINT, g-sequences, and navigation patterns

15Interestingness measures: Support (hits) and confidence

16

Aggregated Logs, queries, and query results

17

The basic idea of the WUM algorithm

18MINT can express 3 types of constraints (“predicates“)

19

The WUM gseqm algorithm

(B predicates)

20Agenda







21

“What makes people happy?” – a corpus-based approach to

finding happiness

22

Bayes‘ formula and its use for classification

1. Joint probabilities and conditional probabilities: basics P(A & B) = P(A|B) * P(B) = P(B|A) * P(A) P(A|B) = ( P(B|A) * P(A) ) / P(B) (Bayes´ formula) P(A) : prior probability of A (a hypothesis, e.g. that an object belongs to a

certain class) P(A|B) : posterior probability of A (given the evidence B)

2. Estimation: Estimate P(A) by the frequency of A in the training set (i.e., the number of A

instances divided by the total number of instances) Estimate P(B|A) by the frequency of B within the class-A instances (i.e., the

number of A instances that have B divided by the total number of class-A instances)

3. Decision rule for classifying an instance: If there are two possible hypotheses/classes (A and ~A), choose the one that is

more probable given the evidence (~A is „not A“) If P(A|B) > P(~A|B), choose A The denominators are equal If ( P(B|A) * P(A) ) > ( P(B|~A) * P(~A) ), choose A

23

Simplifications and Naive Bayes

4. Simplify by setting the priors equal (i.e., by using as many instances of class A as of class ~A)

If P(B|A) > P(B|~A), choose A

5. More than one kind of evidence

General formula:

P(A | B1 & B2 ) = P(A & B1 & B2 ) / P(B1 & B2) = P(B1 & B2 | A) * P(A) / P(B1 & B2) = P(B1 | B2 & A) * P(B2 | A) * P(A) / P(B1 & B2)

Enter the „naive“ assumption: B1 and B2 are independent given A

P(A | B1 & B2 ) = P(B1|A) * P(B2|A) * P(A) / P(B1 & B2)

By reasoning as in 3. and 4. above, the last two terms can be omitted

If (P(B1|A) * P(B2|A) ) > (P(B1|~A) * P(B2|~A) ), choose A

The generalization to n kinds of evidence is straightforward.

These kinds of evidence are often called features in machine learning.

24

Example: Texts as bags of words

Common representations of texts

Set: can contain each element (word) at most once

Bag (aka multiset): can contain each word multiple times (most common representation used in text mining)

Hypotheses and evidence

A = The blog is a happy blog, the email is a spam email, etc.

~A = The blog is a sad blog, the email is a proper email, etc.

Bi refers to the ith word occurring in the whole corpus of texts

Estimation for the bag-of-words representation:

Example estimation of P(B1|A) :

number of occurrences of the first word in all happy blogs, divided by the total number of words in happy blogs (etc.)

25

WEKA – NaiveBayes and NaiveBayesMultinomial

The WEKA classifier learning scheme NaiveBayesMultinomial implements this model of „the probability that a word occurs in a document given that the document is in that classs“.

Its output is a table giving these probabilities

The WEKA classifier learning scheme NaiveBayes assumes that the attributes are normally distributed.

Needed when the attributes are numerical and not necessarily 0 | 1 Its output describes the parameters of these normal distributions Explanation of the annotations of the attributes:

http://grb.mnsu.edu/grbts/doc/manual/Naive_Bayes.html

Explanation of the error measures: http://grb.mnsu.edu/grbts/doc/manual/Error_Measurements.html#sec:error

26

The „happiness factor“ of Mihalcea & Liu (2006)

“Starting with the features identified as important by the Naïve Bayes classifier (a threshold of 0.3 was used in the feature selection process), we selected all those features that had a total corpus frequency higher than 150, and consequently calculate the happiness factor of a word as the ratio between the number of occurrences in the happy blogposts and the total frequency in the corpus.”

What is the relation to the Naïve Bayes estimators?

27Agenda







28Clustering by information contained in the objects to be clustered (here: documents contain text) – www.kartoo.com

29

The basic idea of clustering: group similar things

Group 1Group 2

Attribute 1

Att

rib

ute

2

Based on http://rakaposhi.eas.asu.edu/cse494/notes/f02-clustering.ppt

30

Idea and Applications

Clustering is the process of grouping a set of physical or abstract objects into classes of similar objects.

It is also called unsupervised learning.

It is a common and important task that finds many applications.

Applications in text analysis / Web content mining, e.g. for search engines:

Structuring search results

Suggesting related pages

Automatic directory construction/update

Finding near identical/duplicate pages

Applications in Web usage mining

Customer/user segmentation

User segmentation for recommender systems / personalization


31Concepts in Clustering

“Defining distance between points Cosine distance (which you already know)

Overlap distance

A good clustering is one where (Intra-cluster distance) the sum of distances between objects in the

same cluster are minimized,

(Inter-cluster distance) while the distances between different clusters are maximized

Objective to minimize: F(Intra,Inter)

Clusters can be evaluated with “internal” as well as “external” measures

Internal measures are related to the inter/intra cluster distance

External measures are related to how representative are the current clusters to “true” classes

– See entropy and F-measure

||

||

RQ

RQ


32

Inter/Intra Cluster Distances

Intra-cluster distance

(Sum/Min/Max/Avg) the (absolute/squared) distance between

- All pairs of points in the cluster OR

- Between the centroid and all points in the cluster OR

- Between the “medoid” and all points in the cluster

Inter-cluster distance

Sum the (squared) distance between all pairs of clusters

Where distance between two clusters is defined as:

- distance between their centroids/medoids

- (Spherical clusters)

- Distance between the closest pair of points belonging to the clusters

- (Chain shaped clusters)

From http://rakaposhi.eas.asu.edu/cse494/notes/f02-clustering.ppt

33

How hard is clustering?

One idea is to consider all possible clusterings, and pick the one that has best inter and intra cluster distance properties

Suppose we are given n points, and would like to cluster them into k-clusters

How many possible clusterings?

!k

k n

• Too hard to do it brute force or optimally• Solution: Iterative optimization algorithms

– Start with a clustering, iteratively improve it (eg. K-means)


34

Classical clustering methods

Partitioning methods

k-Means (and EM), k-Medoids

Hierarchical methods

agglomerative, divisive, BIRCH

Model-based clustering methods


35

K-means

Works when we know k, the number of clusters we want to find

Idea:

Randomly pick k points as the “centroids” of the k clusters

Loop: For each point, put the point in the cluster to whose centroid it is

closest

Recompute the cluster centroids

Repeat loop (until there is no change in clusters between two consecutive iterations.)

Iterative improvement of the objective function: Sum of the squared distance from each point to the centroid of its cluster


36

K Means Example (K=2) For a more complex simulation, see

http://www.cs.tu-bs.de/rob/lehre/bv/Kmeans/Kmeans.html Pick seeds

Reassign clusters

Compute centroids

xx

Reasssign clusters

xx xx Compute centroids

Reassign clusters

Converged!


37

A map of documents, grouped by their topics

38

DocumentAtlas: A two-step procedure

1. Latent semantic indexing: Project documents into a semantic space (dimensionality reduction and identification of commonalities even if vocabulary is different)

2. Multidimensional scaling: Project that space into 2D, preserving the distances as well as possible

Input: a set of documents

Output: a „document map“

39Agenda







40Clustering by information contained in the objects to be clustered (here: documents contain text) – www.kartoo.com

41Clustering by information associated with the objects to be clustered (here: photos are associated with tags) – www.flickr.com

42Clustering by information associated with the objects to be clustered (here: queries are associated with document texts) – (1)

43Clustering by information associated with the objects to be clustered ... (2) – Baeza-Yates, Query Mining, ECIR 2005

1. Create instances of past ( query – result set ) combinations

2. Cluster them by the textual similarity of the (viewed) result documents

3. Use this to recommend a better / an additional query

Result set 1Query 1

Result set 2Query 2

New user

recommend

44

Ranking by similarity and popularity: Examples

45Agenda







46

Internet users are worried about their privacy ...

(results from a meta-study of 30 questionnaire-based studies [TK03])

(results from a meta-study of 30 questionnaire-based studies [TK03])

47

... but are they really?An online shop with a difference

[Berendt, Günther, & Spiekermann, Communications of the ACM,2005]

48

Privacy-related behaviour

Shopping for jackets

Shopping for cameras

[Berendt, Data Mining and Knowledge Discovery, 2002], [Berendt, Postproc. WebKDD 2002]

49

Finding: People are willing to exchange privacy for personalization benefits

Users would provide, in return for personalized content, information on their name (88%), education (88%), age (86%), hobbies (83%), salary (59%), or credit card number (13%).

27% of Internet users think tracking allows the site to provide information tailored to specific users.

73% of online users find it useful if site remembers basic information such as name and address.

People are willing to give information to receive a personalized online experience: 51% or 40%, depending on the study.

[TK03]

50User-centric evaluation: An experimental investigation of the effect of explaining the personalization-privacy tradeoff

[KT05] compared the effects of traditional privacy statements with that of a contextualized explanation on users’ willingness to answer questions about themselves and their (product) preferences.

In the contextualized-explanation condition, participants

answered 8.3% more questions (gave at least one answer) (p<0.001),

gave 19.6% more answers (p<0.001),

purchased 33% more often (p<0.07) ,

stated that their data had helped the Web store to select better books (p<0.035) – even though the recommendations were static and identical for both groups.

[KT05] compared the effects of traditional privacy statements with that of a contextualized explanation on users’ willingness to answer questions about themselves and their (product) preferences.

In the contextualized-explanation condition, participants

answered 8.3% more questions (gave at least one answer) (p<0.001),

gave 19.6% more answers (p<0.001),

purchased 33% more often (p<0.07) ,

stated that their data had helped the Web store to select better books (p<0.035) – even though the recommendations were static and identical for both groups.

(screenshot from Teltzrow, M. & Kobsa, A. (2004). Communication of Privacy and Personalization in E-Business. In Proceedings of the Workshop “WHOLES: A Multiple View of Individual Privacy in a Networked World”, Stockholm, Sweden.

51

But what is privacy?Is it only about data protection?

Phillips, D.J. 2004. “Privacy Policy and PETs: The Influence of Policy Regimes on the Development and Social Implications of Privacy Enhancing Technologies.” New Media & Society 6(6): 691-706

freedom from intrusion

construction of the public/private divide

separation of identities

protection from surveillance (the right to choose belonging)

52Also: whose privacy? Stakeholders and privacy interests:a (partially) fictitious example

users of the system:

passengers

system administrators

Other stakeholders:

airport administration

airport security

airlines

duty-free shop

53

Different privacy interests of the different stakeholders

54Agenda







55

In the preparation of a log file(recommendations for open-source tools are shown in green)

1. Use qualitative methods for application understanding (read!)

2. Inspect the site and the URLs for data understanding

1. Generate Analog reports for getting base statistics of usage

2. Build concept system / hierarchy and mapping: URLs concepts (notation: WUMprep regex)

3. Use WUMprep for data preparation

1. Remove unwanted entries (pictures etc.)

2. Sessionize

3. Remove robots

4. Replace URLs by concepts

5. (Build a database)

4. Use WEKA for modelling

1. Transform log file into ARFF (WUMprep4WEKA)

2. Cluster, classify, find association rules, ...

5. Use WUM for modelling

6. Select patterns based on objective interestingness measures (support, confidence, lift, ...) and on subjective interestingness measures (unexpected? Application-relevant?)

7. Present results in tabular, textual and graphical form (use Excel, ...)

8. Interpret the results

9. Make recommendations for site improvement etc.

56

In the case study:

1. Use qualitative methods for application understanding (read!)

2. Inspect the site and the URLs for data understanding

1. Generate Analog reports for getting base statistics of usage

2. Build concept system / hierarchy and mapping: URLs concepts (notation: WUMprep regex)

3. Use WUMprep for data preparation

1. Remove unwanted entries (pictures etc.)

2. Sessionize

3. Remove robots

4. Replace URLs by concepts

5. (Build a database)

4. Use WEKA for modelling

1. Transform log file into ARFF (WUMprep4WEKA)

2. Cluster, classify, find association rules, ...

5. Use WUM for modelling

6. Select patterns based on objective interestingness measures (support, confidence, lift, ...) and on subjective interestingness measures (unexpected? Application-relevant?)

7. Present results in tabular, textual and graphical form (use Excel, ...)

8. Interpret the results

9. Make recommendations for site improvement etc.

done

57

The preparation of texts(e.g., for an automatic version of step 2.2.)

Is quite involved when done properly

(a good introduction to preprocessing for text mining can be found in

Grobelnik, M., & Mladenic, D. Text Mining Tutorial.

http://eprints.pascal-network.org/archive/00000017/01/Tutorial_Marko.pdf )

However, as a first step, you can also use the raw text of documents (generated with only a few of the tools in the TextGarden library).

58

Thank you!

1 Web Usage Mining Modelling: frequent-pattern mining I (sequence mining with WUM), classification and clustering) Prof. Dr. Bettina Berendt Humboldt Univ.

Documents