The Use of Query Reformulation to Predict Future User Actions
Download as PPT, PDF1 like540 views
Using query reformulation data from search logs, researchers can build profiles of users that describe what they are interested in, what they may do in the future, and other attributes. Specifically, analysis of query sequences can provide information about a user's location, topics of interest, complexity of searches, commercial or purchase intent, and likelihood of clicking on results. While analysis of individual queries provides some insights, examining large datasets is needed to develop automated methods for comprehensive user profiling.
![Using Query Reformulation for User Profiling Jim Jansen College of Information Sciences and Technology The Pennsylvania State University [email_address] Interested in how much descriptive information we can generate about a people by leveraging search log data .](https://image.slidesharecdn.com/jjansenqueryreformulationasist09-091111214123-phpapp01/85/The-Use-of-Query-Reformulation-to-Predict-Future-User-Actions-1-320.jpg)
![How much can we tell from a single query? Leveraging demographic information (cont’d): 32.3 percent probability this user has a doctorate 53 percent likelihood works in academia. Using IP , we can locate the geographical area Based on time , could infer that: this person is searching for the conference’s schedule (if the query is submitted prior to the meeting) for travel or looking for presentations or papers from the meeting (if the query is submitted after the conference). Theoretically, we can tell a lot ! However, with billions of queries per month, we can’t do the analysis by hand like this example. To develop user profiles, we need automated methods . Research Question - How complete of a profile can one develop for a Web search engine user from search log data? [(a) what the user is doing, (b) what the user is interested in, and (c) what the user intends to do]](https://image.slidesharecdn.com/jjansenqueryreformulationasist09-091111214123-phpapp01/85/The-Use-of-Query-Reformulation-to-Predict-Future-User-Actions-9-320.jpg)
![Thank you! (open for questions and further discussion) Jim Jansen College of Information Sciences and Technology The Pennsylvania State University [email_address]](https://image.slidesharecdn.com/jjansenqueryreformulationasist09-091111214123-phpapp01/85/The-Use-of-Query-Reformulation-to-Predict-Future-User-Actions-18-320.jpg)
The Use of Query Reformulation to Predict Future User Actions
- 1. Using Query Reformulation for User Profiling Jim Jansen College of Information Sciences and Technology The Pennsylvania State University [email_address] Interested in how much descriptive information we can generate about a people by leveraging search log data .
- 2. What Did We Find Out? We can tell quite a lot about a user! When combined with other information, query reformulation is a revealing searching characteristic.
- 3. The State of Web Search Why search data is important
- 4. The Power of Search and the Web Search is the top online activity Search drives over 7 billion monthly queries in the U.S. Online activity has a huge impact on people’s daily lives: 70 minutes less with family 30 minutes less TV 8.5 minutes less sleep Sources: comScore, U.S., Feb. ’06, Stanford Institute for the Quantitative Study of Society, Nov. ‘05
- 5. Analysis of Search Marketplace Holding fairly stable over the last year or so, albeit with some Bing flux
- 6. Search Logs Contains the trace data recorded when a person visits the search engine, submits a query, views results, etc On one hand, logs have been criticized for not being rich enough (i.e., only have behaviors but not the ‘why ’ factors) On the other hand, logs have been criticized for recording too much about us (i.e., logging a lot of personal information about a person) search logs How much we can learn about a person from the data stored in search logs? Specifically, how rich of a searcher profile can we build of what a person is doing, of why they are doing it, and to predict what are they going to do next?
- 8. How much can we tell from a single query? ASIS&T is an acronym for the American Society of Information Science and Technology Good probability that this user is an academic , a researcher, a librarian, or a student in one of these disciplines Leveraging demographic information : 57 percent female / 43 percent male probability 66.2 percent chance works in the information science field 55.6 percent probability this user has master’s degree
- 9. How much can we tell from a single query? Leveraging demographic information (cont’d): 32.3 percent probability this user has a doctorate 53 percent likelihood works in academia. Using IP , we can locate the geographical area Based on time , could infer that: this person is searching for the conference’s schedule (if the query is submitted prior to the meeting) for travel or looking for presentations or papers from the meeting (if the query is submitted after the conference). Theoretically, we can tell a lot ! However, with billions of queries per month, we can’t do the analysis by hand like this example. To develop user profiles, we need automated methods . Research Question - How complete of a profile can one develop for a Web search engine user from search log data? [(a) what the user is doing, (b) what the user is interested in, and (c) what the user intends to do]
- 10. Specific aspects with automated methods … Location Geographical interest Topical interest Topical complexity Content desires Commercial intent Purchase intent Potential to click on a link Gender User identification – where the user is at – where the user is going – what the user is interested in – how motivated is the user – Info, Nav, Transactional – eCommerce related – getting ready to buy – will user click on link - demographic targeting/personalization - specific user targeting
- 11. Automated methods using query reformulation Location Geographical interest Topical interest Topical complexity – n-grams pattern analysis Content desires Commercial intent Purchase intent Potential to click on a link Gender User identification
- 12. Where to get full story? The methodological implementation reported in paper in your ASIST proceedings: Jansen, B.J., Zhang, M., Booth, B. Park, D., Zhang, Y., Kathuria, A. and Bonner, P. (2009) To What Degree Can Log Data Profile a Web Searcher? Proceedings of the American Society for Information Science and Technology 2009 Annual Meeting. Vancouver, British Columbia. 6-11 November.
- 13. Topical Complexity Number of queries by a user in a session on a topic can tell us many things: the complexity of the topic the user’s motivation for the need provide prediction of future action
- 14. Information Searching Probabilistic user modeling increasingly important area allows computer systems to adapt to users Algorithmic techniques typically employ state models Simple Bayesian Classifier Markov Modeling n-grams Note: not always ‘informational’ anymore. Many time people are searching for ‘ other things ’. Rose & Levinson (2004); Jansen, Booth, & Spink (2008).
- 15. Illustration of Probabilistic User Modeling Using n-grams Given these states … … how accurately can we predict these? AC 5 A 4 ABCDE 3 ABCDE 2 ABCF 1 Search State Transitions User 40% D C 60% B A 100% E CD 66% D BC 1OO% C AB Accuracy Next State? Predictive Pattern
- 16. Example Using Search Log ~ 965,000 searching sessions ~ 1,500,000 queries 8 states focusing on query reformulation Similar results for other aspects of searching See - Qui (1993), Jansen (2005), Jansen & McNeese (2006) Maybe ‘states’ are not the correct paradigm? Jansen, B. J., Booth, D. L., & Spink, A. (2009). Patterns of query modification during Web searching . Journal of the American Society for Information Science and Technology . 10% improvement from 1 st to 2 nd order: okay, but would like to do better 0 1 st 2 nd 3 rd 4 th Order of the Model Accuracy of Prediction 0.1 0.2 0.3 0.4 0.5 0.6 0.28 0.40 0.47 0.44 0.44 0.60 Drop out rate (folks who don’t submit a query ~40%)
- 17. User Profiling Framework Classify user aspects into two levels: internal and external . Internal aspects refer to attributes of the users themselves. External aspects relate to the behavior or interest of the users. Interaction between internal and external aspects. Can infer external aspects from internal aspects. External aspects reflect internal aspects
- 18. Thank you! (open for questions and further discussion) Jim Jansen College of Information Sciences and Technology The Pennsylvania State University [email_address]
- 19. Search Logs has some common fields, such as time, queries, results, etc. We can enrich the log with additional fields. Back Back
- 20. Back
- 21. Back