AIOTB: Benefiting Graph Contrastive Learning from Adversarial Information Optimal Transport Bottlenecks

Graph contrastive learning (GCL) is a widely used graph representation learning strategy, applied to many realworld tasks. When learning a graph representation model, existing GCL methods primarily consider contrastive losses on the graph embeddings while paying little attention to regularizing the node embeddings of each graph. In this study, we propose a novel regularizer called Adversarial Information Optimal Transport Bottleneck (AIOTB), which leads to an effective regularized GCL method. Given a graph, our method randomly samples two subgraphs as its two augmented views and derives their node embeddings layer by layer using a graph representation model. Guided by the information bottleneck principle, AIOTB maximizes the maximal mutual information between same-layer embeddings across views while minimizing the mutual information between adjacent layers within each view, thus producing representations that are cross-view consistent and progressively compressed. We instantiate this information bottleneck based on optimal transport distance and then learn the model in an adversarial learning framework. Experiments demonstrate that incorporating AIOTB consistently benefits state-of-the-art GCL methods across various graph learning tasks.