A Study Of Click Models In Web Search

Search engine has become one of main entries to the Internet. A majority of Internetusers use search engine to seek for information. As to search engines, the quality of searchresults is very important to user experience. Therefore, search engine companies usuallycollect various user interaction data, e.g. query log and click log, to help improve searchquality. Based on these implicit feedback data, click models are widely used to mine therelevance information. It tries to model user clicks by making assumptions on how usersexamine and click on search result pages and then estimate the relevance parameters.The existing click models have considered factors that may impact user clicks, such asposition bias and user satisfaction. In this work, we believe that some other factors havenot been considered yet which may also impact user clicks. For example, user-specificfactors, query-specific factors and time-sensitive factors. We will focus on studying theinfluence of users’ behavioral preference, users’ search expertise and query type. Thenwe build new click models with these factors incorporated.User preference: we study users’ examination behavior when they are searchingthrough eye-tracking experiment. The results turned out that users have diferent exam-ination preferences, indicated by the big variance of users’ average examination depth.Besides, we also analyzed a real search engine click log and found that users behavediferently with respect to click position and click amount. According to these findings,we propose a click model framework which incorporates user preferences and the exper-imental results on multiple click models showed improved performance.User search expertise: click is usually regarded as the signal of user judging adocument as relevant. However, we believe that users have diferent probabilities ofmaking the right relevance judgment facing a document. We give rise to the notion ofuser search expertise, which is assumed to reflect how well the user makes relevancejudgment. Based on this idea, we build click models with user expertise factor embedded.The experimental results showed improved performance on relevance inference with thenew models.Query type: we find that users behave diferently on diferent query types by eye-tracking experiments. The existing click models do not explicitly take query type factorinto consideration. After studying the influence of query type factor to users’ examination behavior, click behavior and search expertise, we propose a click model framework thatincorporates query type factor in multiple ways. This framework allows us to learn querytype information from click-through data and click-based query features in an unsuper-vised manner. It estimates parameters for each query type, for which reason the modelperformance gets improved. Besides, the estimated parameters are consistent with thecorresponding ones found in our independent eye-tracking experiment, which implicitlyvalidates the efectiveness of our method.