Temporal recommendation via modeling dynamic interests with inverted-U-curves

Abstract

How to capture user interest accurately to enhance the user experience is a great practical challenge in recommender systems. Through preliminary investigation, we find that each user has his personalized interest model which may contain multiple kinds of interests, and the strength of each user interest usually has a dynamic evolution process which can be divided into two stages: rising stage and declining stage. The evolution rate of the user interests also differ from each other. Based on this finding, a recommendation framework called SimIUC is proposed, which can identify multiple user interests and adapt the inverted-U-curve to model the dynamic evolution process of user interests. Specifically, SimIUC differs from the traditional user preference based methods which use monotonously decreasing function to model user interest. It can predict the evolutionary trends of interests and make recommendations by inverted-U-interest-based collaborative filtering. We studied a large subset of data from MovieLens and netflix.com respectively. The experimental results show that our method can significantly improve the accuracy in recommendation.