Label-invariantAugmentationforSemi-Supervised GraphClassification

Neural Information Processing Systems 

We conjecture the current performance of graph contrastive learning might be limited by the violation of the label-invariant augmentationassumption.