You Need to Look Globally: Discovering Representative Topology Structures to Enhance Graph Neural Network

Abstract

Graph Neural Networks (GNNs) have shown great ability in modeling graph-structured data. However, most current models aggregate information from the local neighborhoods of a node. They may fail to explicitly encode global structure distribution patterns or efficiently model long-range dependencies in the graphs; while global information is very helpful for learning better representations. In particular, local information propagation would become less useful when low-degree nodes have limited neighborhoods, or unlabeled nodes are far away from labeled nodes, which cannot propagate label information to them. Therefore, we propose a new framework GSM-GNN to adaptively combine local and global information to enhance the performance of GNNs. Concretely, it automatically learns representative global topology structures from the graph and stores them in the memory cells, which can be plugged into all existing GNN models to help propagate global information and augment representation learning of GNNs. In addition, these topology structures are expected to contain both feature and graph structure information, and they can represent important and different characteristics of graphs. We conduct experiments on 7 real-world datasets, and the results demonstrate the effectiveness of the proposed framework for node classification.