AdversarialGraphAugmentationtoImprove GraphContrastiveLearning

Graph contrastivelearning (GCL), by training GNNs to maximize the correspondence between the representations of the same graph in its different augmented forms, may yield robust and transferable GNNs even without using labels.