Graph Contrastive Learning with Generative Adversarial Network

Abstract

Graph Neural Networks (GNNs) have demonstrated promising results on\nexploiting node representations for many downstream tasks through supervised\nend-to-end training. To deal with the widespread label scarcity issue in\nreal-world applications, Graph Contrastive Learning (GCL) is leveraged to train\nGNNs with limited or even no labels by maximizing the mutual information\nbetween nodes in its augmented views generated from the original graph.\nHowever, the distribution of graphs remains unconsidered in view generation,\nresulting in the ignorance of unseen edges in most existing literature, which\nis empirically shown to be able to improve GCL's performance in our\nexperiments. To this end, we propose to incorporate graph generative\nadversarial networks (GANs) to learn the distribution of views for GCL, in\norder to i) automatically capture the characteristic of graphs for\naugmentations, and ii) jointly train the graph GAN model and the GCL model.\nSpecifically, we present GACN, a novel Generative Adversarial Contrastive\nlearning Network for graph representation learning. GACN develops a view\ngenerator and a view discriminator to generate augmented views automatically in\nan adversarial style. Then, GACN leverages these views to train a GNN encoder\nwith two carefully designed self-supervised learning losses, including the\ngraph contrastive loss and the Bayesian personalized ranking Loss. Furthermore,\nwe design an optimization framework to train all GACN modules jointly.\nExtensive experiments on seven real-world datasets show that GACN is able to\ngenerate high-quality augmented views for GCL and is superior to twelve\nstate-of-the-art baseline methods. Noticeably, our proposed GACN surprisingly\ndiscovers that the generated views in data augmentation finally conform to the\nwell-known preferential attachment rule in online networks.\n