1. Characterizing Web Content , User Interest, and Search Behavior by Reading Level and Topic
Date: 2012/11/15
Author: Jin Young Kim, Kevyn Collins-Thompson,
Paul Bennett and Susan Dumais
Source: WSDM 2012
Advisor: Jia-ling Koh
Speaker: Chen-Yu Huang

2. outline Introduction Reading Level & Topic Profiles
Characterizing the web
Applications
Conclusion

3. Introduction
web search
user interest
Topic
Reading level

4. Introduction
Estimate probabilistic profiles to describe users, queries or websites and analyze user behavior
Topic
Reading level

5. outline Introduction Reading Level & Topic Profiles
Characterizing the web
Applications
Conclusion

6. Reading Level & Topic Profiles
entity ：website(s), user(u), query(q)
reading level(R), topic(T), reading level and topic(RT)
profile ：a probability distribution of reading level and topic
(RLT profile) EX
a reading level and topic profile of a user：P(RT | u)
a reading level and topic profile of a query：P(RT | q)

7. Predicting Reading Level and Topic for URL
Represent the reading difficulty of a document as a random variable Rd taking values in the range
Reading Level Classifier
Based on language model
Topic Classifier
Training using URLs in each Open Directory Project category (ODP)

8. Building Reading Level and Topic Profiles
Profiles based on the entity itself
Given a sets of URLs associated with each entity, the joint of distribution of reading level and topic is built by aggregating the distributions of the individual URLs computed by URL-level classifiers
To prevent the bias arising
Choose 25 URLs to estimate the site-level or user-level profiles
Use the top URLs as of the profile for the query

9. Building Reading Level and Topic Profiles
Profiles based on the entity relationships
Circular dependency using profiles based only on the entity itself
Query
Surface
Issue
Website
User
Visit

10. Characterizing and Comparing profiles
Characterizing an Individual Entity
E[R|e] ： expectation of reading level for a given entity e
H(R|e) ： reading level entropy of the entity e

11. Characterizing and Comparing profiles
Characterizing a Group of Entities
Build the profile of an entity group by aggregating the distributions of individual weighted centroid of the individual distributions
EX：reading level profile of U
Characterize the group profile can represent the diversity in terms of its members

12. Characterizing and Comparing profiles
Comparing Entities and Groups
Simplest metric of comparison

13. Characterizing and Comparing profiles
Comparing Entities and Groups
Similarity between the full probability distribution of two entities
Kullback-Leibler(KL) Divergence
Jensen-Shannon(JS) Divergence

14. outline Introduction Reading Level & Topic Profiles
Characterizing the web
Applications
Conclusion

15. Data Set Session Log Data Web content dataset
Contain the anonymized logs of URL visited by user
Web pages visits from users who visited at least 25 pages
During 10 weeks (2010.8)
Web content dataset
Reading level and ODP topic predictions
8 billion web document from

16. Characterizing web content

17. Characterizing websites
Topic-specific analysis

18. Characterizing web queries

19. Characterizing websites
Joint analysis of reading level and topic

20. Characterizing web users
Users’ Deviation from Their Own Profiles
Stretch reading
Future work

21. outline Introduction Reading Level & Topic Profiles
Characterizing the web
Applications
Conclusion

22. Application
Compare expert v.s non-expert URLs

23. Application
Predict expert websites
Result

24. outline Introduction Reading Level & Topic Profiles
Characterizing the web
Applications
Conclusion

25. Conclusion
Provide novel characterizations for websites, users and queries by combining distribution of reading level and topic.
Can be used for a variety of search-related tasks and predicting the content of a URL or site is targeted at domain experts or non-experts.
Use features derived from RLT profiles to predict a user’s preference for Websites in search results. .

26. Conclusion
The divergence metrics developed in this paper can be evaluated for their effectiveness as features for personalized re- ranking.
The techniques developed for expert v.s notice site classification can be applied both for recommendation and ranking purposes.

27. ~Thank you for your listening~