Inferring Web Page Relevance from Human-Computer Interaction Logging

Quality of search engine results often do not meet user’s expectations. In this paper we propose to implicitly infer visitors feedbacks from the actions they perform while reading a web document. In particular, we propose a new model to interpret mouse cursor actions, such as scrolling, movement, text selection, while reading web documents, aiming to infer a relevance value indicating how the user found the document useful for his/her search purposes. We have implemented the proposed model through light-weight components, which can be easily installed within major web browsers as a plug-in. The components capture mouse cursor actions without spoiling user browsing activities, which enabled us to easily collect experimental data to validate the proposed model. The experimental results demonstrate that the proposed model is able to predict user feedbacks with an acceptable level of accuracy.