Collaborative Filtering with Entity Similarity Regularization in Heterogeneous Information Networks

Researchers have been studying hybrid recommender systems which combine user-item rating data with external information in recent years. Some studies suggest that by leveraging additional user and / or item relations, e.g., social network, the performance of the recommendation models can be improved. These studies, nevertheless, mostly utilize a single type of external relationship. Considering the heterogeneity of real-world applications, we propose to position the well-studied recommendation problem in a heterogeneous information network context and attempt to incorporate different recommendation factors. We discuss how heterogeneous information network can benefit recommender systems and then propose a matrix factorization based unified recommendation model to take advantage of both rating data and the related information network. Empirical studies show that our approach outperforms several state-of-the-art recommendation methods on explicit rating data.