Multi-hop Self-augmented Graph Contrastive Learning for Node Classification

Abstract

Current Graph Contrastive Learning (GCL) methods primarily focus on adapting data augmentation techniques from Computer Vision (CV) or Natural Language Processing (NLP) domains. These techniques typically involve modifying input data via node sampling, edge perturbation, or graph structure perturbation. Alternatively, they may adjust the contrastive loss function by increasing or decreasing positive and negative samples based on graph properties. However, few GCL methods have explored designing and discussing the structure of Graph Neural Networks (GNNs) within GCL, despite the significant impact different GNN structures can have on self-supervised GCL performance. Motivated by this gap in research, our paper differs from the approach of most previous methods, designing a Multi-hop Self-augmented GCL method (MSGCL) based on the inherent structural characteristics of GNNs. This method leverages the intrinsic properties of the GNN model structure, utilizing multi-hop information for self-augmentation to generate enhanced views. The approach is simple yet effective, preserving the original graph structure information without resorting to complex and potentially unstable graph structure augmentation methods. We validate this method across five datasets under different labeling conditions. The experimental results indicate that our method surpasses the other advanced methods, even when only a limited number of labels are available. This suggests potential widespread applications in the field of graph signal processing.