Reproduce “SimGRACE: A Simple Framework for Graph Contrastive Learning without Data Augmentation” on OpenResearch

About this paper

Jun Xia, Lirong Wu, Jintao Chen, Bozhen Hu, Stan Z.Li

SimGRACE, from Westlake University, introduces a graph contrastive learning framework that eliminates the need for data augmentation by generating correlated views through perturbed GNN encoders. It achieves competitive or superior performance across various graph learning benchmarks while being 40-90 times faster than some state-of-the-art methods and improves adversarial robustness through AT-SimGRACE.

Key insights

• Perturbing GNN encoder weights directly can generate effective positive pairs for contrastive learning, preserving data semantics more effectively than traditional graph augmentations.
• The effectiveness of SimGRACE stems from its ability to improve both the alignment of positive pairs and the uniformity of representations on the hypersphere.
• Adversarial training on encoder weights, as in AT-SimGRACE, guides the model to learn parameters residing in flatter regions of the loss landscape, theoretically enhancing adversarial robustness.

Summary by alphaXiv.