Deep Graph Contrastive Learning: GRACE and GCA

Introduction Graph representation learning aims to embed nodes or whole graphs into low‑dimensional vectors that capture structural and attribute information. Traditional supervised GNNs face label scarcity and limited transferability, motivating self‑supervised approaches that use proxy tasks.

Contrastive Learning Framework The widely used SimCLR framework consists of three components: random data augmentations, an encoder f followed by a representation extractor g, and a contrastive loss that pulls together representations of augmented views of the same sample while pushing apart different samples. In graphs, augmentations include edge perturbations and feature masking.

Graph Contrastive Learning Methods Representative works include Deep Graph Infomax (DGI), Multiview Graph Contrastive Learning (MVGCL), and Graph Contrastive Coding (GCC). These methods differ in how they generate positive and negative pairs and whether they use global‑local or local‑local objectives.

GRACE Extends SimCLR to graphs by defining intra‑view and inter‑view negative pairs and employing random edge deletion and feature masking as augmentations.

GCA (Adaptive Augmentation) Improves GRACE by assigning removal probabilities to edges and masking probabilities to features based on node centrality, thereby preserving important structural and attribute information.

Experiments Evaluated on Wiki‑CS, Amazon‑Computers, Amazon‑Photo, Coauthor‑CS, and Coauthor‑Physics using node classification accuracy. GRACE and GCA outperform baselines (DeepWalk, node2vec, GAE, VGAE, GraphSAGE, DGI, GMI, MVGRL, GCN, GAT) and demonstrate the benefit of adaptive augmentation.

Conclusions Local‑local contrastive objectives and centrality‑aware augmentations improve unsupervised graph learning, narrowing the gap with supervised methods. Future work should explore better contrastive objectives, data augmentation strategies, and theoretical foundations for graph self‑supervision.