Transcript Automatic Identification of User Interest For Personalized Search

Automatic Identification of
User Interest
For Personalized Search
Feng Qiu, Junghoo Cho
UCLA
Appeared in WWW 2006
Motivation
• Personalized search
– A real-estate agent may issue the query
“office” to look for a vacant office space;
– An IT specialist looks for Microsoft Office!
• Users are reluctant to provide any explicit
feedback or their personal information
Proposed approach
• Learn user’s interest based on user’s click
history
• What’s done in this paper
– Learn users’ interest from click history
– Learn the ranking of user
PageRank
• Key idea
– A page u links to a page v because the author
of page u thinks page v is important.
– If a page links to a lot of pages, the
importance score it confers to each of them
are decreased
• Surf Model
– E(i) is the probability to jump to page i when
she gets bored and n is the total number of
web pages (E(i) = 1/n)
Topic-Sensitive PageRank
Intro to the notations
Topic-Driven Random Surfer Model
• The user browse the web in two-step
process
– Choose a topic of interest t with Probability T(t)
– Jumps to one of the pages on topic t
Topic-Driven Random Surfer Model
(cont.)
Topic-Driven Random Surfer Model
(cont.)
Problem formulation
• Get T
– Linear Regression
– Maximum Likelihood
Rank search results using
preference vector
Evaluation
• Accuracy of topic preference vector
• Accuracy of personalized ranking
– use the accuracy of the final personalized
ranking (as opposed to the accuracy of the
user’s topic preference vector) as our
evaluation metric.
Experimental data
Evaluation (topic preference vector)
Evaluation (personalized ranking)
Appendix