Graph Contrastive Learning with Adversarial Structure Refinement (GCL-ASR)

The scarcity of labeled data in graph neural networks (GNNs) has driven the development of graph contrastive learning (GCL), which has become the most widely used method in unsupervised representation learning. At present, many GCL approaches in the literature utilize random dropout nodes, edges, and masking features to enhance the two views and subsequently enhance consistency between these views. Although this approach is commendable, it fails to address issues such as unequal edge distribution and the varying sensitivity of node training in different augmented views. Given the complex nature of graph data structures, this paper proposes a new GCL framework called Graph Contrastive Learning with Adversarial Structure Refinement (ASR), which explicitly accounts for the graph topology. Specifically, we employ Generative Adversarial Networks (GANs) as a view augmentation strategy to generate high-quality views. Additionally, we introduce a novel training method that incorporates node sensitivity as a regularization term to optimize the GCL training process. We conduct extensive experiments on eight benchmark datasets for node classification, and the results demonstrate that our method has achieved outstanding performance.