Refining conflict in graph contrastive learning via pseudo-labels

Graph contrastive learning (GCL) has emerged as a powerful self-supervised paradigm for graph representation. A critical challenge in GCL arises from the conflict between the contrastive loss function and the message-passing mechanism of graph convolutional network encoders, as the message-passing mechanism pulls neighboring nodes close, while the contrastive loss function pushes negative nodes apart, especially neighboring nodes. Prevalent methods typically tackle the conflict by identifying conflicting node pairs and subsequently ignoring them during training. However, the identification of conflicting pairs in these methods may be one-sided and even incorrect, likely deleting valuable pairs which are inherently non-conflicting. To overcome the limitation, we have proposed conflict refined graph contrastive learning (CR-GCL), a novel framework that leverages pseudo-labels to prevent ignoring non-conflicting negative pairs. Technically, CR-GCL incorporates three key components: 1) a perturbation augmentation module that creates augmentation graphs and generates representations; 2) a conflict quantification module that estimates the gradient impact of negative pairs by tracking representation similarity changes; 3) a conflict refinement module that utilizes pseudo-labels to enhance the identification of conflicting negative pairs, which are then ignored during training. Extensive experiments on six benchmark datasets revealed that CR-GCL significantly outperforms state-of-the-art methods, delivering superior node classification accuracy with various label rates.