MGRec: Multi-Graph Fusion for Recommendation

The rapid development of the Internet and social networks deepens the connection between people and bring more rich information. To solve the problems of information overload, many recommendation algorithms have flourished, such as collaborative filtering algorithms, social recommendation algorithms, and meta-path based algorithms. The common characteristic of these algorithms is that they introduce some additional information, e.g., user friendship or item category, to improve the accuracy of recommendations. However, how to fuse such different types of information effectively remains challenging: how to aggregate specific information within each type of relations and how to combine various information among different types of relations. To solve the aforementioned problems, we propose a graph-level information fusion recommendation algorithm, i.e., MGRec (Multi-Graph Recommendation). Firstly, it models three types of relations, that are, ratings between users and items, friendships between users, and categories between items, as sub-graphs. Then, it makes information aggregation within each sub-graph, using personalized feature extraction methods like GCN and attention mechanism. Finally, it makes information combination among sub-graphs, which is implemented by a gate mechanism. Extensive experiments are conducted on two real-world datasets and results demonstrate the irreplaceability of MGRec.