Incorporating Temporal Information for Recommendation in Heterogeneous Information Network

Abstract

Heterogeneous information network is a better representation of various types of data, which has been successively applied to many fields. It provides a promising way to deal with data sparsity problem and improves the quality of recommendation. But current HIN-based recommendations ignore the time attributes on the relationships and fail to capture the dynamic changes of users' preference. In this paper, we use the temporal information to enhance the recommendation quality in HIN. Firstly, we propose the concepts of time-weighted HIN and extend the relevance measurement of time-weighted meta-path by designing a time deviation matrix. Then, SVD matrix factorization is applied to generate the latent features of each meta-path. To better assemble recommendations from different meta-paths, we propose a new weight learning approach called complementary combination (CC). Experiments on Movielens100K dataset illustrate the improvement of our work.