Pairwise Uncertainty-Based Graph Contrastive Learning

Abstract

Currently, vanilla graph contrastive learning (GCL) has become the mainstream paradigm for learning discriminative node representations in unsupervised settings. However, its performance is limited by two issues. First, traditional graph augmentation strategies lack awareness of edge semantic relationships, which may cause uncertain signals to propagate along the graph structure. Second, the model pushes away all negative samples equally, ignoring the rich discriminative information embedded in highly uncertain negative pairs. To address these issues, we propose a pairwise uncertainty-based GCL method named PUGCL. Specifically, PUGCL divides affinity samples and estrangement samples for each anchor by fitting similarity distributions, and an uncertainty estimator is trained in a pairwise manner. Subsequently, the estimated uncertainty is used to guide the design of graph augmentation and contrastive loss. In this process, PUGCL adopts a differentiated treatment. Edges with high uncertainty are removed to purify the information propagation signal, while negative pairs with high uncertainty are assigned higher weights to refine the decision boundary. Extensive experiments on node classification and node clustering across six datasets are conducted, and the results demonstrate that PUGCL outperforms state-of-the-art methods.