 Who Uses Web Search for What? And How?

Contribution  Combine behavioral observation and demographic features of users  Provide important insight on search behavior  Design decision  Search results  Information flows

Key Ideas  Who is searching  Analyzing user’s demographics  What are they searching for  Analyzing query topics  How they are searching  Analyzing session information

User Modeling  Query topics (what?): what are the topics that user issues queries on?  Y! Directory classification for topics  User demographics (who?): What is the demographics profile of the user?  Used a mix of user-provided information (age and gender) and information derived from user’s zip code  Session characteristics (how?) : Does a user have many/few, short/long and navigational/information sessions?  Session length  No. of queries per session

Data  web search query log of Yahoo! Search engine (2.3 million)  Log from U.S. Yahoo! site  Registered U.S. Yahoo users with an identifiable cookie  Only active users 1. Issue at least 100 queries over the sample period 2. Remove users with more than 100,000 queries issues 3. Remove users who clicked fewer than 1/100 of their queries 4. Remove User who clicked more than 100 time per queries

“Who?” Data  Using U.S Census with zip code Per-capita income Level of education Ethnicity  Using 2008 U.S. presidential election with zip code to obtain result

“Who?” Data

“What?” Data  Topic distribution of the queries issued by user  Use top 10 Yahoo! search results obtained for a given query  Use unique proprietary classification to classify into 71 topics  Classified queries issued at least 30 times  10 million distinct queries

“Who?” Data

“How?” Data  Session length and no. of clicks per session  Count no. of queries issued within session interval of 30 mins  Classified queries into Navigational- seek a single website Informational - queries that cover a board topic Transactional- queries that reflect the intent of user to perform a particular action (Click Entropy) H(D|q) = ∑ d,q –p(d|q) log 2 p(d|q)  Compute click entropy for queries issued >= 20 times  H(D|q) ≤ 1 – focused query  H(D|q) ≥ 3 – diverse query

“How?” Data  370k distinct queries for both focused and diverse queries  Focused queries – navigational  Diverse queries - informational

Method  Unsupervised K-mean clustering  K ranging from 8 to 20  Clustered users using topics distributions  Using the result to induced the “Who” and “How”

Informational Users  Used search engine as research engine to find information on a wide range of topics  “Who” Well educated with above-average income  “What” Do research on a wide range of topics with little interest in adult content  “How” More likely to issue non-navigational queries Less likely to have a single-click session More likely to make use of the suggested query alternatives

Navigational Users  Used search engine as a replacement for web page bookmarking to navigate to URLs that he already knows exist  “Who” Background averages of topical cluster under consideration  “What” Dominated by topic of popular website e.g FB, gmail  “How” More likely to issue navigational queries More likely to click only on a single result within a session Less likely to make use of unnecessary suggested query alternatives

Transactional Users  Used search engine to take him to some URL where he can perform desired transaction Little benefit in learning more about a subject URL generally not known in advance  “Who” Depend heavily on the kind of transaction  “What” Predominant topics are shopping, adult content and gaming sites  “How” Diverse clicks

Close-Up  Baby Boomers Avg age of 50 yrs old Simple navigational queries related to online banking and interested in finance  Liberal Females Most likely to voted for Democrat in 2008 elections Biggest single query is shopping with longer session  White Conservatives Voted for Republican in 2008 elections Search for automotive related topics, business pages & home and garden information

Conclusion  Overall finding are stereotypical  Future work Fine grained analysis in term of categories and search strategy Closer look at long-tail queries