Using viewing time to infer user preference in recommender systems

Abstract

The need for effective technologies to help Web users locate items (information or products) is increasing as the amount of information on the Web grows. Collaborative filtering is one of the most successful techniques for making recommendations; however, most CF-based systems require explicit user ratings and a large quantity of usage history to function effectively. In addition, such systems typically rely on comparing a user to ‘similar ’ users encountered before. We develop and evaluate the idea that viewing time is an indicator of preference for attributes of items, and a recommendation system based on this idea. The system uses only the current user’s navigational data in conjunction with item property data to make recommendations. We also present empirical evidence that the system makes useful recommendations.