FA-GCL: Feature-augmented graph contrastive learning method

Graph contrastive learning seeks to improve the efficacy of graph representation learning by comparing various graph representations. Existing approaches predominantly rely on node attributes or structural information for contrastive analysis. However, in real-world applications, node attribute information can be incomplete or entirely absent, while structure-enhancement methods often generate false positive samples. To mitigate these issues, we propose a feature augmentation-based graph contrastive learning method (FA-GCL) that enhances the accuracy and robustness of graph representations. Specifically, our approach first implements a dynamic dropout-based feature augmentation technique, which adjusts dropout rates dynamically using a triangular wave function, thereby significantly improving model performance. Additionally, we introduce two complementary feature augmentation methods based on singular value decomposition: a theoretically rigorous full SVD approach and a computationally efficient randomized projection-based SVD method that achieves linear complexity while preserving spectral properties. Both methods add controlled noise to singular values and reconstruct features to create high-quality augmented samples. Comprehensive experiments were conducted on twelve widely used graph datasets. The results indicate that FA-GCL consistently outperforms baseline methods in node classification, node clustering, and graph classification tasks.