Spying Out Real User Preferences for Metasearch Engine Personalization

Most current metasearch engines provide uniform service to users but do not cater for the specific needs of individual users. To address this problem, research has been done on personalizing a metasearch engine. An interesting and practical approach is to optimize its ranking function using clickthrough data. However, it is still challenging to infer accurate user preferences from the clickthrough data. In this paper, we propose a novel learning technique called “Spy Näıve Bayes” (SpyNB) to identify the user preference pairs generated from clickthrough data. We then employ ranking SVM to build a metasearch engine optimizer. To evaluate the effectiveness of SpyNB on ranking quality, we develop a metasearch engine prototype that comprises three underlying search engines: MSNSearch, WiseNut and Overture to conduct experimental evaluation. The empirical results show that, compared with the original ranking, SpyNB can significantly improve the average ranks of users’ click by 20%, while the performance of the existing methods are not satisfactory.