1. Introduction

While the current search engines are already very useful to us, they are far from optimal and all have a fundamental limitation -- they cannot distinguish individual users. When figuring out which web pages are likely interesting to a user, these search engines generally only use the keywords provided by a user in the query; as a result, a car shopper who uses the word "jaguar" to search for information about the Jaguar car would get exactly the same results as a zoologist who might use the same word to find information about the jaguar animal.

The goal of the UCAIR project is to break this limitation and develop the next-generation search engine technologies that can better understand an individual user's information need and optimize search results according to each individual user. Our research includes (1) developing a new UCAIR framework based on Bayesian decision theory; (2) developing new language models to exploit user information and search context to improve retrieval accuracy; (3) developing new retrieval methods to optimize the long-term retrieval utility over an entire retrieval session; (4) developing new retrieval methods to leverage user similarities to better infer one particular user's information need based on information about other similar users; and (5) developing prototype UCAIR systems for searching the Web and bioinformatics literature.

2. Team

• Current members
• ChengXiang Zhai (Professor)
• Alexander Kotov (Ph.D. student)
• Bin Tan (Ph.D. student)
• Xuanhui Wang (Ph.D. student)
• Past members
• Azadeh Shakery (Ph.D., 2008)
• Hui Fang (Ph.D., 2007, Ohio State University)
• Tao Tao (Ph.D., 2007, Microsoft)
• Xuehua Shen (Ph.D., 2007, Google)
• Aditya Ramani (MS, 2006)
• Smitha Sriram (MS, 2003)

• Retrieval framework: Traditional retrieval frameworks are inadequate for modeling and exploiting user information and search context. To support user-centered adaptive information retrieval in a general way, we proposed a general decision theoretic framework for optimizing interactive retrieval [Shen et al. CIKM 05].
• Implicit feedback: A user's search history, including all the past queries and clickthrough information, contains valuable clues about the user's interests and search preferences. We have developed several algorithms to use such history information to improve search accuracy for a current query of the user. These algorithms have been shown to be effective for both short-term implicit feedback [Shen et al. SIGIR 05a] and long-term implicit feedback [Tan et al. KDD 06]. Since no extra user effort is required, these implicit feedback techniques can be applied to any search engine.
• Explicit feedback: When a user is willing to provide explicit feedback, it is important to maximize the usefulness of feedback information from the user so that the user can use minimum effort to obtain maximum benefit. We have developed active feedback algorithms to optimize the documents to be presented to users for relevance feedback, which are shown to be better than presenting top-k results [Shen et al. SIGIR 05b]. We have also developed term feedback techniques that can improve retrieval accuracy through obtaining user feedback at the level of terms [Tan et al. SIGIR 07].
• Negative feedback: When the initial search results are extremely poor, all the top-ranked documents may be non-relevant. To help users in such a situation, we studied how to learn from the top-ranked non-relevant documents (i.e., negative feedback) to improve the ranking of the unseen documents and proposed effective methods for negative feedback [Wang et al. SIGIR 08].
• Personalized organization of search results: The search history of users can also be used to improve organization of search results. In [Wang et al. SIGIR 07], we proposed techniques to learn interesting aspects of user information needs from search logs and use such knowledge to cluster search results in a user-oriented way.
• UCAIR toolbar: We have developed an intelligent software agent (i.e., the UCAIR toolbar), which was shown to outperform Google by up to 20% in precision in a user study [Shen et al. CIKM 05]. As an "information broker" owned by a user, the software agent would observe a user's search behavior very closely, analyze all the queries from the user and all the web pages viewed by the user, and use such extra user information to customize and improve search results from a search engine for the user. UCAIR toolbar is available for download at http://sifaka.cs.uiuc.edu/ir/proj/ucair/download.html .

  The picture on the right shows how UCAIR can re-rank search results from Google and optimize search results for a user searching information about the Jaguar car using the query "jaguar". The left side shows the original mixed results with pages about Jaguar cars and Jaguar software. The right side shows the automatically re-ranked results by UCAIR after the user has viewed the 2nd page, which is about Jaguar car. The new results no longer have pages about the Jaguar software; instead, two new pages about Jaguar cars have been pushed up by UCAIR, which were originally ranked down in the results from Google.

4. Selected Publications (See all publications)

Xuanhui Wang, Hui Fang, ChengXiang Zhai. A study of methods for negative relevance feedback , Proceedings of the 31st Annual International ACM SIGIR Conference on Research and Development in Information Retrieval ( SIGIR'08 ), to appear. ( 17% acceptance)

5. Funding Support

• National Science Foundation, CAREER grant IIS-0347933
• Google Research Grant
• Microsoft Live Labs Research Grant
• UIUC Faculty Startup

[ Team ] [ Results ] [ Publications ] [ Funding ]

---

Text Information Management Group Home | About| People | Projects| Publications | Downloads| Links| Wiki