ISIRS: information theory-based social influence with recommender system

Abstract

With explosive growth of information, recommender systems have made great progress during past ten years. The improvement in accuracy of recommendation has great commercial value. However, the accuracy still has a large space to improve, and the cold-start problem also restricts the performance of recommender systems. Aiming at optimising these two problems, ISIRS model is proposed. ISIRS integrates social influence into recommendations. Considering celebrity effect in sociology, ISIRS applies information theory to capture the social influence in a social network. As a result, ISIRS can find famous persons in a social network by sorting social influence of all people. ISIRS then makes use of the preferences of these famous people to make recommendations more accurate. The results of experiments show that ISIRS model outperforms the recommendation based on users, the recommendation based on items and the MF recommendation algorithm, even though the rating matrix and trust relationship are sparse. These results prove ISIRS can help both the accuracy and the cold-start problem in recommendations.