Invariant Risk Minimization Augmentation for Graph Contrastive Learning

Despite significant advancements in Graph Contrastive Learning (GCL) in recent years, effective and interpretable graph data augmentation methods remain a challenge in the field. Traditional graph augmentation methods typically involve randomly modifying node attributes, edge properties, or extracting subgraphs, with their fundamental principles and operations derived from random processing of image data. However, graph data lacks the Euclidean structure and attribute intuitiveness of image data, making it difficult to ensure that augmentation data retains key features. In this paper, we propose a novel method called Invariant Risk Minimization Augmentation for Graph Contrastive Learning (IRMGCL). IRMGCL incorporates invariant risk minimization loss during the training process to ensure that data augmentation methods preserve invariant features. Specifically, we use different augmentation channels as environmental variables and train these channels to maintain invariant features through different environments. Experiments conducted on citation network datasets Cora and Citeseer demonstrate the effectiveness of our method. We compared our method with state-of-the-art graph contrastive learning methods, showing superior performance. Simultaneously, we conducted ablation experiments to verify the role of each component loss function in IRMGCL. Furthermore, our framework can incorporate labeled data for semi-supervised contrastive learning. Our experiments show that model performance can be further enhanced.