TechFest SIS poster

1
Analysis of Topic Dynamics in Web Search
Xuehua Shen (Univ of Illinios)
Susan Dumais, Eric Horvitz (Microsoft
Research)

Topics
Arts, Business, Computers, Games, Health, Home,
KidsTeens, News, Recreation, Reference,
Science, Shopping, Society, Sports, Adult
Log Data
Topic Dynamics

• Queries (ID, Time, QueryString, Topic)
• Clicks (ID, Time, URLVisited, Topic)
• Detailed analysis of 6k users
• 100 actions during first two weeks
• 660k total URL visits
• Automatically assigned Topics to URLs

• Key Characteristics
• What topics do users look at?
• How consistent are user topics over time?
• How similar are users to other users, groups of
  users, or the population at large?

Results
Modeling

• Predict topics visited based on history
• Using data from
• Individuals
• Groups of similar individuals
• Population as a whole

Groups Models

• Markov gt marginal
• Group best

Time Lag

• Using models of
• Marginal
• Base probability for each topic
• Markov
• Probability of moving from one topic to another
• Time-Specific Markov
• Time-dependent Markov models

• Small lags best

Time-Specific Models

• Short time gaps best

Research Directions

• Reliability and failure modes of automatic topic
  tagging
• Analysis of different user groups
• Combining query and click through information
• Improved personalization of the search experience