Multi-Granularity Contrastive Learning for Graph with Hierarchical Pooling

Peishuo Liu, Cangqi Zhou, Xiao Liu, Jing Zhang, Qianmu Li

Abstract: Graph contrastive learning is an unsupervised learning method for graph data. It aims to learn useful representations by maximizing the similarity between similar instances and minimizing the similarity between dissimilar instances. Despite the success of the existing GCL methods, they generally overlook the hierarchical structures of graphs. This structure is inherent in graph data and can facilitate the organization and management of graphs, such as social networks. Therefore, the representation results learned from previous methods often lack important hierarchical information in the graph, resulting in suboptimal performance for downstream tasks. In this paper, we propose a \(\textbf{M}\)ulti-\(\textbf{G}\)ranularity \(\textbf{G}\)raph \(\textbf{C}\)ontrastive \(\textbf{L}\)earning (\(\textbf{MG2CL}\)) framework that considers the hierarchical structures of graphs in contrastive learning. This method enables effective learning of better graph representations by combining view information at different resolutions. In addition, we add a cross-granularity contrast module to further improve the accuracy of representations. Extensive experiments are conducted on seven graph classification datasets to demonstrate the effectiveness of MG2CL in learning unsupervised graph representations.