Link prediction in recommender systems based on multi-factor network modeling and community detection

Abstract

Link prediction provides methods to estimate potential connections in complex networks, which has theoretical and practical significance for personalized recommendation and various other applications. Traditional collaborative filtering and other similar approaches have not utilized sufficient information on the community structure of networks. Therefore, this paper presents a link prediction model based on complex network modeling and community detection. In the approach, complex networks are constructed by considering the similarity among users' preference for genre selection, the similarity among users' rating distribution, and the similarity among items based on users' ratings. And the similarity calculation results are taken as weight of links as well as objects are considered as nodes in networks. On this basis, the community detection results can be obtained, and link prediction is performed with the community information considered. Multi-factor community detection based on node similarity improves the prediction process effectively and increases accuracy in our experiments. The result infers that users' behaviors, including rating an item and selecting an item over others, indicate a hidden community structure in the system, which can be used for link prediction and even for better understanding of complex systems.