ECDL 2006 - keynote 2 - Yahoo determining intentionality from queries

11:00 Keynote 2

Queries and Clicks as a Source of Knowledge

Ricardo Baeza-Yates
Yahoo Research
Barcelona, Spain & Santiago, Chile

http://www.dcc.uchile.cl/~rbaeza/

web (chaotic) - DL (ordered)

[overview of Yahoo sites, amount of info, types of data]

"Information Games" - win if you match tags
http://www.espgames.org/

Observed Data

* Query Logs
* Result/Web Clicks
* Advertising clicks
* Social

Talks about "safe" (trusted) vs "dangerous" (false, spammy) information/sources.

The Wisdom of Crowds

The Power of Social Media

Motivations for Web Mining

* The Dream of the Semantic Web
  - Obstacle: Us

* User Actions: Implicit Semantic Information
  - free
  - large volume
  - unbiased
  - can we capture it?
  - hypothesis: Queries are the best source

Mining Queries for...

* Improved Web Search
* User Driven Design
  - Information Scent
  - web site that users want
  - web site hat you should have
  - imprve content & structure
* Bootstrap of pseudo-semantic resources

Web Queries

* short queries & impatient interaction
* smaller and different vocab
* different user goals (Broder, 2000)
  - information need
  - navigational need
  - transactional need
* Refined by Rose & Levinson, WWW 2004

http://citeseer.ist.psu.edu/rose04understanding.html

Yahoo Mindset

http://mindset.research.yahoo.com/

Relevance of the Context

* moving to less information, more context

Context

* Who you are
* Where you are
* What you are doing

* Issues: privacy ...
* Sources: Web, CV, usage logs...
* Goals: personalization, localization, better ranking in general...

Context in Web Queries

* IP, time, location (based on IP), interaction history, task, OS, browser...

User Intention

* Kang & Kim, SIGIR 2003
  - their method was not effective: 60%

* Liu, Lee & Cho WWW 2005
  - prediction power 90%

Yahoo

* Manual classification of more than 6000 popular queries
  - query intention and topic
  - classification and clusering
  - machine learning

* Baeza-Yates ? 2006?

Results: You can do (machine learning?) classification of intention on information queries

Next step: Clustering Queries

* Define relations among querys
  - common words
  - common clicked URLs: works better = natural clusters
* define distance functon among queries

Yahoo Approach

* Can we cluster queries well?
* Can we assign user goals to queries?

[details of method]

The user queries represent in a way the user view of your data/system.

Uses

* Improved ranking
* Word classification
  - e.g. synonyms in the same cluster
* Query recommendation (ranking of suggested queries)

Building Taxonomies

* Infer topics for queries that imply documents

Result: Automatic classification is better than (single or a small group of) humans! (but/because the auto classification is based on the actions of many many people)

Final Remarks

* Many potential uses of the wisdom of people

Q: Compare user pseudo-taxonomies to taxonomies autogenerated from text
A: probably they are quite different - people use different words (or a different majority set of words) in queries than they do in written text

Q (Rachel Heery UKOLN): can this pick up new terms (new words) being used?
A: working on this