CAREER: Improving Information Access by Learning from User Interactions

点击翻译按钮获取中文摘要

The project takes a machine learning approach to improving the effectiveness of information access tools, in particular the retrieval quality of search engines. The ability to learn enables a search engine to automatically adapt its retrieval strategy to individual users, to specific user groups, and to particular WWW sites. A search engine should learn, for example, that a query for ``Michael Jordan'' issued from a user at cs.cornell.edu is much more likely to refer to the professor at UC Berkeley than for an average user. Similarly, a search engine should be able to adapt to collection properties, for example, that in a particular intranet not the TITLE field, but the H1 headlines contain the most important information.Since explicit user feedback is rarely available, implicit feedback derived from observable user behavior is used as the input to the learning algorithms. Such implicit feedback requires new machine learning methods, since it comes in forms that are different from the standard machine learning settings. For examples, in search engines it is more reasonable to exploit clickthrough data as feedback in the form of pair-wise preferences (e.g. ``for query Q, document A should be ranked higher than document B'') than as an absolute relevance feedback. The project analyzes the reliability of implicit clickthrough data, designs and analyzes learning methods, and evaluates their applicability on an educational database, providing a service to the scientific community. Beyond this direct contribution, this technology can be used to improve the performance of general purpose search engines such as Google and hence has broader impacts beyond the scientific community. Information on this project is available on the web http://www.cs.cornell.edu/People/tj/career. The resulting software will be made available for download.