Web search has become a fundamental means for massive users to find information. As a result, huge amounts of user interaction data are generated and act as a valuable source for many web tasks. An important task is to understand the user preference of each query-document pair based on their click behavior, so as to allow search engines to deliver better search results to effectively serve their users. In this thesis, we study the problem of modeling user click behavior in Web search, which is often formalized as a click model problem. Click models can automatically infer user-perceived relevance for each search result. This in turn enforces the search engines to deliver better search results....[ Read more ]

Web search has become a fundamental means for massive users to find information. As a result, huge amounts of user interaction data are generated and act as a valuable source for many web tasks. An important task is to understand the user preference of each query-document pair based on their click behavior, so as to allow search engines to deliver better search results to effectively serve their users. In this thesis, we study the problem of modeling user click behavior in Web search, which is often formalized as a click model problem. Click models can automatically infer user-perceived relevance for each search result. This in turn enforces the search engines to deliver better search results.

In the context of a click, there are multiple objects: user, query, session, task, search result, page region and so on. Previous models generally treat each object in isolation, disregarding their associations by only considering individual queries and search results. This may bring an over-simplification to a model but sacrifice valuable associative information. The main contribution of this thesis is a family of models and algorithms to address these limitations via modeling the associations between these objects. The proposed model and algorithm family characterizes the associations from six facets. We first put forward a whole-page model to describe the interplay between organic search and sponsored search. We then propose a session-based model and an intent-bias model to collectively study multiple queries with their corresponding clicks. We then introduce a user-based model to enrich the query and document with the user and characterize this triple relationship. We continue with a novel noise-aware model to capture the noise of a click by leveraging the above objects as its context. Finally, we provide a new solution by combining multiple proposed click models together to solve the relevance prediction challenge. We further verify all the proposed models through extensive experiments using large-scale data collected from a commercial search engine. Experimental results demonstrate the significant improvements over the state-of-the-art.