SAIL: Self-Augmented Graph Contrastive Learning

This paper studies learning node representations with graph neural networks (GNNs) for unsupervised scenario. Specifically, we derive a theoretical analysis and provide an empirical demonstration about the non-steady performance of GNNs over different datasets, when supervision signals are not appropriately defined. The depends on both feature smoothness locality structure. To smooth discrepancy proximity measured by topology feature, proposed SAIL - novel self-augmented contrastive framework, two complementary self-distilling regularization modules, i.e., intra- inter-graph knowledge distillation. We demonstrate competitive variety applications. Even single GNN layer, has consistently or even better various benchmark comparing state-of-the-art baselines.