IRC: An Iterative Reinforcement Categorization Algorithm for Interrelated Web Objects

1
IRC An Iterative Reinforcement Categorization
Algorithm for Interrelated Web Objects

• Gui-Rong Xue, Dou Shen, Qiang Yang
• Hua-Jun Zeng, Zheng Chen, Yong Yu and Wei-Ying Ma

2
Outline

• Motivation
• Related Work
• Categorization of interrelated objects
• Experiments
• Conclusion

3
Document Classification

• Definition and Applications
• Approaches
• SVM
• Naïve Bayes
• kNN
• Document Representation
• VSM model
• N-gram model

4
Webpage Classification

• Difference to Document Classification
• Page semi-structure
• Hyperlink and anchor text
• Challenges
• Noisy content ads, navigation bar, images,
  scripts
• No coherent page-construction style, language or
  structure
• New Approaches
• Link-based classifications
• Summarization based classification

5
Heterogeneous Web
Query
Query Thesaurus
Query Session Log
Query Log
Hyperlink
How to leverage the inter-relationship between
the heterogeneous data objects?
Web-page
Browsing patterns
Human Relationship
User Profile
6
Virtual-document

• One possible approach virtual-document
• To augment the features of Web objects by using
  their interrelated Web objects as additional
  features
• A little improvement over the content-based
  categorization approach
• There is still head room for harnessing such
  interrelationships between the heterogeneous
  objects.

Original Feature vector
Additional Feature Vector
Virtual Feature Vector
7
Our Approach

• Iterative Reinforcement Categorization (IRC)
• The category information of one object is
  reinforced by the category information of all its
  interrelated objects
• The updated category information of the object
  will consequently reinforce the category
  information of its interrelated objects
• Stop until it converges to a conclusive result.
• To fully exploit the relationships over the
  heterogeneous data objects.

8
Outline

• Motivation
• Related Work
• Categorization of interrelated objects
• Experiments
• Conclusion

9
Related Work

• Classification on the Web pages
• Content-based
• Joachims 13 proposed a method of using Support
  Vector Machines (SVMs) to classify documents.
• Dumais and Chen 7 use text representation to
  organize search results into an existing
  hierarchical structure.

10
Related Work

• Classification on the Web pages
• Link-analysis-based classification
• Handle both text components of the Web pages and
  the hyperlink relationship.
• Cohn et al. 6 and Glover et al.9 combining
  link-based and content-based.
• Chakrabati et al. 2 Probabilistic model.
• Getoor et al. 8 PRMS
• Our work is different
• We could classify the heterogeneous data objects
  across different data types.
• The data objects of different types could not be
  taken as in the same space.

11
Related Work

• Query log analysis
• Beeferman and Berger 1 Iterative clustering on
  clickthrough data.
• Ignore the content features of the query, the Web
  page.
• Wen et al. 19 described a query clustering
  method.
• Wang et al. 17 proposed a method of using query
  clickthrough log to iteratively reinforce query
  and Web page clusters.

12
Outline

• Motivation
• Related Work
• Categorization of interrelated objects
• Experiments
• Conclusion

13
Categorization of interrelated objects

• Problem Definition
• Iterative reinforcement algorithm

14
Categorization of interrelated objects Problem
Definition

• The Web could be modeled as a weighted directed
  bipartite graph G(V, E)
• the nodes in V represent queries and Web pages,
• the edges E represent the clickthrough
  information from a query to a browsed Web page.
• In this paper, we divide V into two subsets
• Queries Qq1, q2, , qm
• Web pages Dd1, d2, , dn
• A matrix M is used to represent the adjacency
  matrix, whose (i, j)-element is the weight from
  Web page i to query j.

15
Categorization of interrelated objects Problem
Definition

• Given a bipartite graph G(V, E), the problem is
  how to classify the Web pages D and queries Q
  into a set of predefined categories Cc1, c2, ,
  ck, where k is the number of categories.

C1 C2 C3 Ck
Web pages
Queries
Web pages
Queries
Web pages
Queries
16
Categorization of interrelated objectsIterative
reinforcement algorithm

• Basic Idea
• We classify the Web pages according to their
  content feature,
• To fully utilize the relationship in the
  bipartite graph, we propose a novel iterative
  reinforcement classification method.
• The basic idea is to propagate the categories
  computed for one type of object to all related
  objects by updating their probability of
  belonging to a certain category.
• This process is iteratively performed until the
  classification results for all object types
  converge.

17
Categorization of interrelated objectsIterative
reinforcement algorithm

• We divide the Web pages D into training set Dt
  and testing set Ds.
• We consider all queries as being part of the test
  set Qs

Web pages Dt
Queries Qs
Web pages Ds
18
Categorization of interrelated objectsIterative
reinforcement algorithm
Step 1
Input
Web pages Dt
Text Classifier
Web pages Ds
Categories of Web pages Ds
Categories of Web pages Dt
Queries Qs
Step 3
Iterative Propagation
Inter-relationship
Step 2
Categories of Queries Qs
19
Categorization of interrelated objectsIterative
reinforcement algorithm

• Step 2 Classifying queries based on a bipartite
  graph
• Infer the probability of queries belonging to
  categories according to their interrelated Web
  pages

Testing
Training
20
Categorization of interrelated objectsIterative
reinforcement algorithm

• Step 3 Classifying Web pages based on a
  bipartite graph
• Re-classify Web pages through the relationship
  between queries and pages

Queries Testing
Web pages Content
21
Categorization of interrelated objectsIterative
reinforcement algorithm

• Steps 2-3Iterative reinforcement categorization
• Then,
• After several iterations, the PS and RS would
  reach a fixed point.

22
Outline

• Motivation
• Related Work
• Categorization of interrelated objects
• Experiments
• Conclusion

23
ExperimentsData set

• A set of classified Web pages extracted from the
  Open Directory Project (ODP) (http//dmoz.org/)
• A real MSN query clickthrough log is collected.
• We deal with the common pages which appeared in
  both the ODP data set and the query clickthrough
  log
• 131,788 Web pages in 15 top-level categories
• 199,564 associated queries
• 468,696 relationships between Web pages and
  Queries.

24
ExperimentsFeature selection, Classifier and
Evaluation Criteria

• A simple feature selection method, known as
  Information Gain (IG)20, is applied in our
  experiments
• We take the SVM as the content based classifier
• Evaluated using the conventional precision,
  recall and F1 measures

25
ExperimentsPerformance

• Baseline
• content-based classification method
• Comparison take the interrelated queries could
  be taken as an additional feature for their
  corresponding pages
• Query-metadata based classification
• Only use the query metadata as features of Web
  page directly
• Virtual document based classification
• Web page content and query metadata

26
ExperimentsPerformance

• IRC achieves the higher performance than the
  other three methods in comparison.
• F1-micro-averageing measure
• Over the content method by 26.4,
• Over the query metadata method by 21,
• Over the virtual document method by 16.4.

27
ExperimentsPerformance

• The effect of the clickthrough data by increasing
  the clickthrough data size

28
ExperimentsPerformance

• The page length has an important effect on the
  performance of classification based on content
  feature
• The error rate of the content-based
  categorization gradually increasing along with
  the shorter length of pages
• While our IRC still keeps in a stable quality.

29
ExperimentsPerformance

• The performance of the IRC algorithm with the
  iteration times
• The convergence curve of our iterative algorithm
• The execution time of the algorithm on different
  data size

30
Outline

• Motivation
• Related Work
• Categorization of interrelated objects
• Experiments
• Conclusion

31
Conclusion

• The novelty of our work can be seen from several
  aspects.
• First, we extend the traditional classification
  methods to multi-type interrelated data objects.
• Aim to classify interrelated data objects of
  different types simultaneously using both their
  content features and their relationship with
  other types of objects.
• Second, we present a reinforcement algorithm to
  classify interrelated Web data objects on a
  bipartite graph.
• The category of one type is propagated to
  reinforce the categorization of other
  interrelated data objects, vice versa, as an
  iterative process.

32
Reference

• D. Beeferman and A. Berger. Agglomerative
  clustering of a search engine query log. In
  Proceedings of the sixth ACM SIGKDD International
  Conference on Knowledge Discovery and Data
  Mining, pages 407-415, 2000.
• S. Chakrabarti, B. Dom, and P. Indyk. Enhanced
  hypertext categorization using hyperlinks. In
  Proceedings of the ACM SIGMOD International
  Conference on Management of Data, pages 307-318,
  Seattle, Washington, June 1998.
• C.Cortes and V. Vapnik. Support Vector Networks.
  Machine Learning, 201-25, 1995.
• S. L. Chuang and L. F. Chien. Enriching Web
  taxonomies through subject categorization of
  query terms from search engine logs. Decision
  Support System, Volume 35, Issue 1, April 2003.
• H. Cui, J. R. Wen, J. Y. Nie, and W. Y. Ma. Query
  Expansion by Mining User Logs, IEEE Transaction
  on Knowledge and Data Engineering, Vol. 15, No.
  4, July/August 2003.
• D. Cohn and T. Hofmann. The missing link - a
  probabilistic model of document content and
  hypertext connectivity. In Advances in Neural
  Information Processing Systems 13, pages
  430-436.MIT Press, 2001.
• S. Dumain and H. Chen. Hierarchical
  Classification of Web Content. In Proceedings of
  the 23rd annual international ACM SIGIR
  Conference on Research and Development in
  Information Retrieval, 2000.
• L. Getoor, N. Friedman, D. Koller, and B. Taskar.
  "Learning Probabilistic Models of Relational
  Structure," In Proceeding of the 18th
  International Conference on Machine Learning,
  2001.
• E. J. Glover, K. Tsioutsiouliklis, S. Lawrence,
  D. M. Pennock, and G. W. Flake. Using Web
  structure for classifying and describing Web
  pages. In Proceedings of WWW-02, International
  Conference on the World Wide Web, 2002.
• G. Grimmett and D. Stirzaker. Probability and
  Random Processes, 2nd ed. Oxford, England Oxford
  University Press, 1992.
• C. K. Huang, L. F. Chien, and Y. J. Oyang.
  Relevant term suggestion in interactive Web
  search based on contextual information in query
  session logs. JASIST 54(7) 638-649, 2003.
• G. Jeh and J. Widom. SimRank A Measure of
  Structural-Context Similarity. Proceedings of the
  Eighth ACM SIGKDD International Conference on
  Knowledge Discovery and Data Mining, pages
  538-543, Edmonton, Canada, July 2002.
• T. Joachims. Text categorization with support
  vector machines learning with many relevant
  features. In Proceedings of ECML-98, 10th
  European Conference on Machine Learning, pages
  137-142, Chemnitz, Germany, April 1998.
• H. J. Oh, S. H. Myaeng, and M. H. Lee. A
  practical hypertext categorization method using
  links and incrementally available class
  information. In Proceedings of the 23rd annual
  international ACM SIGIR conference on Research
  and development in information retrieval, pages
  264-271. ACM Press, 2000.
• Sequential Minimal Optimization, http//research.
  micro-soft.com/jplatt/smo.html.
• S. Slattery and M. Craven. Discovering test set
  regularities in relational domains. In
  Proceedings of ICML-00, 17th International
  Conference on Machine Learning, pages 895-902,
  Stanford, US, 2000.
• J. D. Wang, H. J. Zeng, Z. Chen, H. J. Lu, L.
  Tao, and W.-Y Ma. ReCoM reinforcement clustering
  of multi-type interrelated data objects. In
  Proceedings of the ACM SIGIR Conference on
  Research and Development in Information
  Retrieval, pages 274-281, Toronto, CA, July 2003.
• J. Platt. Probabilistic outputs for support
  vector machines and comparisons to regularized
  likelihood methods. In A. Smola, P. Bartlett, B.
  Scholkopf, and D. Schuurmans, editors, Advances
  in Large Margin Classi ers. MIT Press, 1999.
• J. R. Wen, J. Y. Nie, and H. J. Zhang. Clustering
  user queries of a search engine. In Proceedings
  of the Tenth International World Wide Web
  Conference, Hong Kong, May 2001.

33
Thanks!