Study of Effectiveness of Implicit Indicators and Their Optimal Combination for Accurate Inference of Users Interests

Retrieval and filtering systems may apply relevance feedback to gain information on users’ needs in order to improve their ad-hoc queries or long term profiling. Explicit relevance involves explicit ratings of documents or terms by the users and disrupts their normal patterns of browsing and searching. The alternative non-disruptive method is implicit feedback inferring users’ needs and interests by monitoring their regular interaction with the system. Some implicit indicators of interest, such as reading time, have been investigated in previous studies and were found indicative to the relevance of documents but not sufficiently accurate to substitute explicit ratings. In this paper we present several new relative implicit feedback indicators and examine their effectiveness as well as the effect of combining several implicit indicators. The paper describes a large-scale user study on which users’ searches were observed by a specially developed browser that recorded their behavior (implicit indicators) as well as their explicit ratings. The relationship between implicit indicators and explicit ratings was analyzed and found that a certain combination of implicit indicators achieved higher correlation with the explicit ratings than any of the individual indicators. We have also found that the newly suggested relative indicators are more indicative to the level of interest of the user in an information item than the non-relative indicators.