Improving Structural and Semantic Global Knowledge in Graph Contrastive Learning with Distillation

Graph contrastive learning has emerged as a pivotal task in the realm of graph representation learning, with the primary objective of maximizing mutual information between graph-augmented pairs exhibiting similar semantics. However, existing unsupervised graph contrastive learning approaches face a notable limitation in capturing both structural and semantic global information. This issues poses a substantial challenge, as nodes in close geographical proximity do not consistently possess similar features. To tackle this issue, this study introduces a simple framework for Distillation Node and Prototype Graph Contrastive Learning (DNPGCL). The framework enables contrastive learning by harnessing similar knowledge distillation to obtain more valuable structural and semantic global indications. Experimental results demonstrate that DNGCL outperforms existing unsupervised learning methods across a range of diverse graph datasets.