Recommender System could model users' interests based on their historical behaviors in the systems. As the dramatic expanding of Web 2.0 web sites and applications, the information about users and items and the links between them are of different type, which is also called heterogeneous information. Although there are many papers researching on the collaborative filtering and social recommendation system, solid works about how to effectively utilize the heterogeneous information to help improve the recommendation performance are rare. This paper investigate the personalized recommendation on Heterogeneous Information Network, including Top-N recommendation task and rating prediction task using heterogeneous user-item information. Following are the main contributions of this paper: 1. Meta Path-based Top-N recommendation on Heterogeneous Information Network. Meta Path in heterogeneous network links various types of nodes using different types of edges, which could carry different semantic information. We model the heterogeneous relations between users and items using various Meta Paths, which link users with items by heterogeneous nodes and edges. We propose a novel algorithm called Meta Path-based Score Propagation which could efficiently compute how much a user would like the items through a specific kind of Meta Path. Using this score, we can recommend the Top-N graded items to this user. Moreover, we combine different kinds of Meta Paths using supervised learning and then make recommendation based on the hybrid score from these paths. Experimental results on real data sets show the algorithms we propose in this chapter could achieve higher accuracy in Top-N recommendation. 2. We propose two heterogeneous information Graph Regularization-based Matrix Factorization algorithms for the rating prediction task. The two algorithms effectively utilize useful heterogeneous information and the network structure to improve the performance of matrix factorization. Furthermore, the proposed algorithms are more general than previous related works and could extent to different types of heterogeneous information network. We prove the efficiency of the two algorithms using experiments on two real data sets, and our methods could make more accurate prediction compared with other state-of-art algorithms.
修改评论