Contrastive Graph Representation Learning via Maximizing Mutual Information

Abstract

With the expansion of graph data in the real world, unsupervised graph representation learning shows greater potential. Unsupervised graph representation learning is mainly about extracting high-level representation from the rich properties and structures of graphs. In this paper, we propose Graph Representation Contrastive Learning Framework based on mutual information(CgI).It extracts the effective topology structural information and context of the graph through maximizing the node-level and graph-level mutual information in two perspectives respectively. Node-level mutual information mainly focuses on local association information between nodes, while graph-level mutual information is more concerned with the guiding role of global information. CGI combines contrastive learning and mutual information into feature extraction. The experimental results confirm that the proposed model performs outstanding improvements contrast with the-state-ofthe-art models, and it has better representative learning ability.