Self-supervised Graph-level Representation Learning with Adversarial Contrastive Learning

The recently-developed unsupervised graph representation learning approaches apply contrastive into graph-structured data and achieve promising performance. However, these methods mainly focus on augmentation for positive samples, while the negative mining strategies are less explored, leading to sub-optimal To tackle this issue, we propose a Graph Adversarial Contrastive Learning (GraphACL) scheme that learns bank of samples effective self-supervised whole-graph learning. Our GraphACL consists (i) encoding branch generates representations (ii) an adversarial generation produces samples. generate more powerful hard our method minimizes loss during updating maximizing adversarially over providing challenging task. Moreover, quality produced by is enhanced through regularization carefully designed divergence orthogonality loss. We optimize parameters alternately. Extensive experiments fourteen real-world benchmarks both classification transfer tasks demonstrate effectiveness proposed approach existing methods.