{Tian-Ming Bu Fan Zhang}

Abstract

Graph neural networks (GNNs) have been the dominant approaches for graph representation learning. However, most GNNs are applied to homophily graphs and perform poorly on heterophily graphs. Meanwhile, these GNNs fail to directly capture long-range dependencies and complex interactions between 1-hop neighbors when generating node representations by iteratively aggregating directly connected neighbors. In addition, structural patterns, such as motifs which have been established as building blocks for graph structure, contain rich topological and semantical information and are worth studying further. In this paper, we introduce the common-neighbors based motifs, which we called CN-motifs, to generalize and enrich the definition of structural patterns. We group the 1-hop neighbors and construct a high-order graph according to CN-motifs, and propose CN-motifs Perceptive Graph Neural Networks (CNMPGNN), a novel framework which can effectively resolve problems mentioned above. Notably, by making full use of structural patterns, our model achieves the state-of-the-art results on several homophily and heterophily datasets.