Combining Meta-Graph and Attention for Recommendation over Heterogenous Information Network

Recently heterogeneous information network (HIN) has gained wide attention in recommender systems due to its flexibility in modeling rich objects and complex relationships. It’s still challenging for HIN based recommenders to capture high-level structure and fuse the mined features of users and items effectively. In this paper, we propose an approach for the recommendation over HIN, called MGAR, which combines Meta-Graph and Attention to address the challenge. Informally speaking, meta-graph is applied to feature extraction, so as to capture more semantic information, while the attention mechanism is used to fuse the features arising from different meta-graphs. MGAR can be divided into two stages. In the first stage, we apply the matrix factorization technique to generate latent factors based on predefined meta-graphs. In the second stage, the embeddings of users and items are fused with the neural attention mechanism. And then the deep neural network is employed to make recommendations by modeling complicated interactions. Experiments over two real datasets indicate MGAR achieves state-of-the-art performance.