Graph Contrastive Learning with Local and Global Mutual Information Maximization

Abstract

Graph is a common data structure in social networks, citation networks, bio-protein molecules and so on. Recent years, Graph Neural Networks(GNNs) have attracted more and more research attention because of its superior performance on some graph learning tasks. Training of GNNs needs large amount of labeled data, which casts shadows on the usability and expansibility of GNNs. Inspired by recent unsupervised research on computer vision and natural language processing, we propose a novel unsupervised graph representation learning model together with several graph data augmentation methods (drop edge, blur, mask features) and a local and global graph mutual information maximization strategy. By maximize two types of mutual information between original graph and augmented graph, the model is forced to learn some useful prior domain knowledge. We conduct experiments on both node classification and graph classification tasks and show the superior performance of the proposed model over state-of-the-art baselines.