Adaptive Multi-View Joint Contrastive Learning on Graphs

Recently, contrastive learning has shown promising results for representing graphs. Despite their success, several key issues have not been well addressed in existing studies: 1) Data noise and incompleteness are inevitable in the graph signal due to various factors. 2) Improper augmentation strategies may have negative effects on views construction and graph representation. In this study, we propose a novel joint contrastive learning model for graph representation named MVJCL. Specifically, a set of views were constructed with topology-level and node-level augmentation strategies. For each view, we execute two-layers GCNs to learn the node embeddings. Then, we propose a positive-negative-positive (pnp) contrastive learning task, which performs contrastive learning between the negative view and each positive view, so as alleviate the noise of supervision signal and exploit the most critical information. Extensive experiments on five real-world datasets demonstrate the effectiveness of MVJCL, where the maximum improvement can reach to 4.63%.