ImGCL: Revisiting Graph Contrastive Learning on Imbalanced Node Classification

Graph contrastive learning (GCL) has attracted a surge of attention due to its superior performance for node/graph representations without labels. However, in practice, the underlying class distribution unlabeled nodes given graph is usually imbalanced. This highly imbalanced inevitably deteriorates quality learned node GCL. Indeed, we empirically find that most state-of-the-art GCL methods cannot obtain discriminative and exhibit poor on classification. Motivated by this observation, propose principled framework Imbalanced classification (ImGCL), which automatically adaptively balances from Specifically, first introduce online clustering based progressively balanced sampling (PBS) method with theoretical rationale, training sets pseudo-labels obtained We then develop centrality PBS better preserve intrinsic structure graphs, upweighting important graph. Extensive experiments multiple datasets settings demonstrate effectiveness our proposed framework, significantly improves recent methods. Further experimental ablations analyses show ImGCL consistently representation under-represented (tail) classes.