Multiple Neighbor Relation Enhanced Graph Collaborative Filtering

Abstract

Graph convolutional networks (GCNs) have substantially advanced state-of-the-art collaborative filtering (CF) methods. Recent GCN-based CF methods have started to explore potential neighbor relations instead of only focusing on direct user-item interactions. Despite the encouraging progress, they still suffer from two notable limitations: (1) only one type of potential neighbor relations is explored, i.e., co-interacting with the same item/user, neglecting the fact that user-item interactions are associated with various attributes and thus there can exist multiple potential neighbor relations from different aspects; (2) the distinction between information from direct user-item interactions and potential neighbor relations and their different extents of influence are not fully considered, which represent very different aspects of a user or an item. In this paper, we propose a novel Multiple Neighbor Relation enhanced method for Graph Collaborative Filtering (MNR-GCF) to address these two limitations. First, in order to capture multiple potential neighbor relations, we introduce a new construction of heterogeneous information networks with multiple types of edges to account for multiple neighbor relations, and a multi-relation aggregation mechanism to effectively integrate relation-aware information. We then enhance CF with a degree-aware dynamic routing mechanism to dynamically and adaptively fuse information from direct user-item interactions and potential neighbor relations at each aggregation layer. Our extensive experimental results show that our solution consistently and substantially outperforms a large number of state-of-the-art CF methods on three public benchmark datasets.