Attentive Meta-graph Embedding for item Recommendation in heterogeneous information networks

Heterogeneous information network (HIN) has become increasingly popular to be exploited in recommender systems, since it contains abundant semantic information to help generate better recommendations. Most conventional work employs meta-paths to model the rich semantics in the HIN. However, the meta-path as a linear structure is insufficient to express the connections. Recently, several work adopts a graph structure, i.e. meta-graph, to express the complex semantics. However, they treat the contributions of nodes in the meta-graph equally, and no explicit representations for users, items or meta-graph based context are learned in the process. To tackle the above problems, this paper proposes an Attentive Meta-graph Embedding approach for item Recommendation, called AMERec, in HINs. Firstly, we prioritize those highly similar pairwise features in the selection of meta-graph instances. Secondly, we differentiate each node in the meta-graph and learn an embedding for each meta-graph. Thirdly, we consider the differences between user and item pairs based on their meta-graph context, and learn a weight for each meta-graph by leveraging the attention mechanism. Finally, we predict the rating by capturing the low- and high-dimensional interaction information between users, items and their meta-graph based context. Comprehensive experiments on three different datasets show that the proposed method is superior to other comparative methods. (C) 2020 Elsevier B.V. All rights reserved.