Layer-Adaptive-Augmentation-Based Graph Contrastive Learning With Feature Decorrelation.

Abstract

Graph Contrastive Learning (GCL) methods typically leverage augmentation techniques to generate different graph views for comparison, thereby learning corresponding representations for graph-related tasks in label-scarce scenarios. However, existing GCL methods suffer from two primary limitations: 1) they use predefined or one-time perturbations for augmentation, ignoring adaptive noise injection during forward propagation and thus leading to suboptimal model robustness; 2) their contrast mechanisms mainly focus on the agreement of inter-graph representations while neglecting the dimensional feature redundancy within intra-graph representations. To solve these issues, we propose Layer-adaptive-augmentation-based Graph Contrastive Learning with feature Decorrelation (LGCLD). First, the designed layer-wise adaptive augmentation method performs dynamic perturbations while maintaining the semantic similarity between augmented and original graphs, which can improve model robustness. Second, we introduce an Agreement-Decorrelation loss (AD loss) that simultaneously optimizes the agreement between graph-level representations and the feature correlation among different dimensions within each graph-level representation, promoting the model to learn informative and non-redundant graph-level representations. Furthermore, we analyze the reasonableness of AD loss through the graph information bottleneck principle. Experiments on various-domain graph datasets demonstrate that LGCLD achieves better or competitive performance compared with a series of state-of-the-art baselines.