New Perspectives on Search Click Modeling
Click modeling aims to interpret the users’ search click data in order to predict their clicking behavior. Existing models can well characterize the position bias of documents and snippets in relation to users’ mainstream click behavior in either organic search block or ads block. Yet, current advances limit their focus solely on position bias, while click modeling possesses the potential as a much wider topic. In this thesis, we propose two directions of extending existing click model works: (1) expanding query-document relevance score with a user dimension, hence personalized click models capturing user intrinsic preferences by matrix and tensor factorization; and (2) using previous click models as a micro layer for each user click out of a macro click chain, which includes search click logs for every click-able block on a whole search result page. Either one of our perspectives on search click modeling produces a general framework that could incorporate existing click models and remains valid for possible future developments on position bias depiction. We verify both models through extensive experiments using large-scale data collected from a real search engine, and their improvements over current models are significant. In addition, our models are very capable of handling challenging problems in the literature, including prediction on rare queries and ads click interpretation, which may offer inspirations for future research.