TMP: Meta-path based Recommendation on Time-Weighted Heterogeneous Information Networks

Abstract

Heterogeneous information network is a new efficient way of representing data in current recommender systems. It provides rich side information that can be added to deal with the data sparsity problem and produces better recommendations. However, current HINs ignore the time influence on relationships, and the widely used meta-path in HIN fails to capture the temporal changes of users' preferences. In this paper, we extend current HIN and meta-path with time attributes through introducing a time deviation matrix, which can distinguish users' past and recent behaviors. Moreover, we propose a time-weighted meta-path-based recommendation method (TMP) to predict the ratings of users on items, which use the matrix factorization idea of Funk-SVD and combine predicting results from different meta-paths through a weight learning method. To optimize the recommendation, we use user and item biases to address those items which are durably popular and those users who have stable preferences. Experimental results on dataset show the effectiveness of our approach.