Dual-channel hard negative sample generation for graph contrastive learning.

Abstract

Graph Contrastive Learning (GCL) has demonstrated remarkable effectiveness in representation learning for graph-structured data. However, most existing GCL methods have two issues regarding the low-quality of negative samples. Firstly, most simple negative samples are semantically invalid and fail to provide discriminative signals. Secondly, false negative samples, which are caused by indiscriminate processing methods, may degrade the overall performance of the model. To this end, we propose a novel method called Dual-Channel Hard Negative Sample Generation for Graph Contrastive Learning (DCGCL). In it, we employ a dual-channel graph generator to apply controlled perturbations to the original adjacency and feature matrices. This process ensures high structural and feature similarity to the original graphs. To differentiate the semantics of the perturbed and original graphs, we introduce a maximizes distribution differences mechanism. This mechanism maximizes the distributional divergence between the two graph types while constraining the consistency of node probability distributions. Moreover, adopts a two-stage training strategy to dynamically inject hard negative samples in the later training phase for learning more generalized feature representations across samples of varying difficulty. Experimental results on diverse downstream tasks validate that DCGCL significantly enhances the performance of GCL models, outperforming state-of-the-art baselines with notable improvements.