Discovery of Indirect Association from Web Usage Data

1
Discovery of Indirect Association from Web Usage
Data
Vipin Kumar Army High Performance Computing
Research Center Department of Computer Science
Engineering University of Minnesota kumar_at_cs.umn.
edu Homepage http//www.cs.umn.edu/kumar
2
Overview
Internet
Web usage data
3
Overview

• Web Usage Mining
• the automatic extraction of non-trivial patterns
  from Web usage data
• Main techniques
• Clustering
• to find groups of users who share similar
  browsing behavior
• Classification
• to categorize Web users according to their past
  access history
• Association
• to determine what are the set of page views
  often accessed together in the same server
  session

4
Application of Web Usage Mining
Source J. Srivastava, R. Cooley, M. Deshpande,
PN Tan, Web Usage Mining Discovery and
Applications of Usage Patterns from Web Data,
SIGKDD Explorations (2000)
5
Taxonomy of Web Usage Mining
Source J. Srivastava, R. Cooley, M. Deshpande,
PN Tan, Web Usage Mining Discovery and
Applications of Usage Patterns from Web Data,
SIGKDD Explorations (2000)
6
Mining Association Patterns in Web Data
Input
Output
/home ? /product ? /product/electronics
Frequent Itemsets
/home ? /shipping ? /shipping/status
Association Rules
/home ? /product ? /product/offer ?
/product/electronics/tv
Sequential Patterns
/home ? /product ? /product/electronics ?
/product/electronics/tv
Association Patterns
Click-streams
7
Mining Association Patterns in Web Data

• Provide the following information
• What are the set of pages frequently accessed
  together by Web users? (frequent itemsets)
• What page will be fetched next? (association
  rules)
• What are the paths frequently traversed by Web
  users? (sequential patterns)
• Web association patterns are useful
• To improve Web site design
• To develop prefetching and Web caching policies.
• To recommend related pages
• To collect business intelligence about behavior
  of Web users

8
Association Mining is Support-based

• Frequent itemset
• combination of pages that have support greater
  than a user-specified minimum threshold
• Association Rule

• Higher support implies greater statistical
  significance
• Support-based pruning constraints the exponential
  complexity and makes the association rule
  computation tractable

9
Can Infrequent Patterns be Interesting?

• If support between A and B is too small, then
  there may be a negative correlation between A and
  B
• P(A,B) lt P(A) P(B)
• If (A ? B) has a high support, then support(A,B)
  will tend to be low because
• P(A B) P(B) P(A,B)
• Example Coffee ? Tea

10
Approach 1 Using Negative Items

• Computationally expensive
• Tends to produce too many negative associations

11
Approach 2 Negative Itemsets

• Savasere et al 1998
• A negative itemset is a set of items whose actual
  support is significantly lower than its expected
  support
• Expected support can be computed using item
  taxonomy

Suppose C and G are frequent
12
Indirect Association
a
M
b
THEN a and b are expected to be frequent
? a and b are indirectly associated via mediator M
13
Non-Sequential Indirect Association Formulation

• A pair of Web pages, a and b, are indirectly
  associated via mediator set M, if
• Sup(a,b) lt ts
• Sup(a ? M) ? tf, Sup(b ? M) ? tf
• Dep(a,M) ? td, Dep(b,M) ? td
• where Dep(x,M) can be any reasonable objective
  measures

14
Finding Interesting Negative Associations
For all pairs of items
With Mediator
No Mediator
FM
FN
Frequent
Minimum itempair support
IM
IN
If
Infrequent
IM/FM IN/FN
then Indirect Association is not surprising
mediator thresholds
15
Finding Interesting Negative Associations
With Mediator
No Mediator

• IM/FM is small
• IM/IN is small
• ? Indirect Association is interesting

Frequent
FM
FN
Infrequent
IN
IM
16
Finding Interesting Negative Association
Indirect Association is interesting when minimum
itempair support threshold is small. But, if
threshold is too low, very few indirect
associations are obtained.
17
Application Market-Basket Analysis

• Substitute items up-sell
• Pavilion PC ? 17 Monitor ? Pavilion multimedia
  PC
• Competing items competitive analysis
• Coke ? Ruffles ? Pepsi
• Complementary items cross-sell
• Tekken 3 ? Playstation Memory Card ? Tomb Raider 2

18
Other Applications Information Retrieval

• Identify synonyms and antonyms
• Identify the different contexts of a queried word
• Useful to group together query results

Union
Trade
Soviet
Worker
19
Other Applications Stock Market
Indirect Association can partition events that
are associated with the movement of a stock price.
20
Sequential Indirect Association Formulation

• A pair of Web pages, a and b, forms a sequential
  indirect association via a mediator sequence, w,
  if the following conditions are satisfied
• Support(a,b) lt ts
• Support(s1) ? tf, Support(s2) ? tf (s1aw or
  wa, s1bw or wb)
• Dependence(a,w) ? td,
• Dependence(b,w) ? td
• w does not contain a and b
• discover groups of users who have different
  interests but share a common traversal path

21
Types of Sequential Indirect Association
Different entry points
Different entry and user interests
Different user interests
Type C (convergence)
Type D (divergence)
Type T (transitive)
22
Clustering vs Indirect Association

• Clustering is another way to discover different
  groups of users
• A. Banerjee and J.Ghosh, Clickstream Clustering
  using Weighted Longest Common Subsequences,
  Workshop on Web Mining (2001)
• TW Yan, M Jacobsen, H Garcia-Molina and U Dayal,
  From User Access Patterns to Dynamic Hypertext
  Linking, Proc of the 5th International World Wide
  Web Conference (1996).
• YJ Fu, K Sandhu, MY Shih, A Generalization-Based
  Approach to Clustering of Web Usage Sessions, Web
  Usage Analysis and User Profiling, LNAI (2001)
• Clustering cannot find distinct groups of users
  who share a similar traversal path since the
  support of the mediator is large (mediator often
  contains navigation pages)
• It is more likely that several indirect
  associations are contained within a single
  cluster, rather than each indirect association
  connects between two separate clusters

23
Indirect Association for Web Data
24
Impact of Site Structure on Associations

• Navigation pages (e.g. home pages and hub pages)
  tend to have higher support than content pages

Source U of Minnesota Computer Science
department Web logs (Jan 1-31, 2001)

• If threshold is too high, most of the patterns
  contain only navigation pages
• If threshold is too low, too many patterns!

25
How Indirect Association helps

• Indirect association groups together patterns
  that have similar substructures
• The common substructures (mediators) often
  contain the navigation pages

Mediator
Navigation pages
26
Grouping Indirect Associations

• Indirect associations can also be grouped
  together into more compact structures if they
  share a common mediator

Check degree of association
27
Viewer for Non-Sequential Indirect Association
Indirect Pairs (dashed line)
List of Mediators
Frequent Pairs (solid line)
Currently Selected Mediator
28
Experimental Setup

• Data sources
• UMN University of Minnesota Web log
• contains 91,443 page views and 34,526 sessions
• EC E-commerce Web log
• Contains 6664 page views and 143,604 sessions
• Steps
• Preprocessing
• identify sessions, convert sessions into
  transactions
• Extract frequent itemsets or sequences
• apply Apriori algorithm to find frequent itemsets
• Apply GSP algorithm to find frequent sequences
• Apply indirect association algorithm
• Merge indirect associations that have common
  mediator

29
Experimental Results (UMN)
Students taking CS as minor subject
Contact information
Prospective graduate students
CS Graduate student association
Students planning to take PhD exam
30
Experimental Results (UMN)
31
Experimental Results (EC)
32
Experimental Results (EC)
33
Experimental Results (EC)
34
Experimental Results (EC)
35
Experimental Results (EC)
36
Experimental Results (EC)
37
Conclusion

• Indirect Association provides an alternative
  approach to capture interesting infrequent
  patterns
• For Web data, indirect association represents the
  distinct interests of Web users who share similar
  traversal path
• Such patterns cannot be easily found using
  standard association and clustering techniques
• Indirect Association can be used to group
  together patterns into more compact structures
• Navigation pages form the mediators

38
References

• PN Tan, V Kumar, J Srivastava, Indirect
  Association Mining Higher Order Dependencies in
  Data, In Proc of the 4th European Conference on
  Principles and Practice of Knowledge Discovery in
  Databases, Lyon, France, Sept 13-16, 632-637
  (2000)
• PN Tan, V Kumar, H Kuno, Using SAS for Mining
  Indirect Associations in Data, In Proc of the
  Western Users of SAS Software Conference (2001)
• PN Tan, V Kumar, Mining Indirect Associations in
  Web Data, In Proc of WebKDD 2001 Mining Log Data
  Across All Customer TouchPoints, August (2001)
• KDD Web Mining workshops WebKDD 1999, WebKDD
  2000, WebKDD 2001
• SIAM Workshop on Web Mining 2001 and 2002