Modeling User Behavior and Attention in Search

Abstract

In Web search, query and click log data are easy to collect but they fail to capture user behaviors that do not lead to clicks. As search engines reach the limits inherent in click data and are hungry for more data in a competitive environment, mining cursor movements, hovering, and scrolling becomes important. This dissertation investigates how remotely collecting rich user interaction data in the form of mouse cursor activity can help researchers understand fundamental human behavior and improve the design of search engines. Specifically, mining cursor activity can improve upon state-of-the-art methods for scoring and ranking search results, and estimating where users are looking without eye-tracking. Descriptive analyses of cursor movements show how users move their cursor when they search to provide signals of relevance and explain reasons for abandoning a search. User models can be used to infer visual attention on the page to identify what content users are looking at, as well as compute the relevance and attractiveness of search results to the user. This implicit feedback given to the search engine can then inform the layout and content presented on the pages, or improve the ranking of search results. This dissertation will demonstrate the following thesis: users' mouse cursor interactions can be collected efficiently on the Web, used to understand users' search behaviors, and can be useful in the design of Web search engines.