Modeling Long-Term User Profile in Collaborative Filtering

Abstract

Collaborative filtering methods are widely accepted and used for item recommendation in various applications and domains. Their simplicity and ability to provide recommendations without the need for specific domain knowledge makes them utilized in both academic and industrial community. However, continuous dynamics of the system makes them vulnerable on different kind of changes, such as the change in user’s preferences or the appearance of new users and items in the system. It is particularly emphasized in user-based collaborative filtering recommendations which are based on both users and its nearest neighbor’s long-term profiles. In this work an approach is presented in order to recognize some of the changes and provide an upgraded model that provides improved performance. Such changes include deviations in user’s mean ratings, changes in neighbors’ similarities, as well as deviations from neighbors’ mean rating. For each of the proposed improvements a set of matching parameters are identified that form a long-term user’s profile. The performance of the proposed models compared to standard user-based collaborative filtering methods is evaluated in terms of prediction accuracy. The experimental results performed on real data set show sound improvement utilizing adjustment of deviations from user’s mean ratings and the neighbor’s similarities, as well as promising improvement with adjustment of neighbor’s mean ratings.