Graph Contrastive Learning: Foundations, Methods, and Recent Advances (GRACE & GCA)

Graph representation learning aims to embed nodes (or whole graphs) into low‑dimensional vectors that capture structural and attribute information, but most graph neural networks rely on supervised training, which suffers from label scarcity and limited transferability.

Self‑supervised learning addresses these issues by defining proxy tasks; contrastive learning, in particular, encourages representations of different augmented views of the same graph to be similar while pushing apart representations of other graphs. The typical pipeline follows SimCLR: random data perturbations, an encoder (often a GNN) followed by a projection head, and a contrastive loss such as InfoNCE.

Early graph contrastive methods (e.g., Deep Graph Infomax, MVGCL, GCC) adopt either global‑local or local‑local objectives. Building on this, the GRACE model applies a SimCLR‑style local‑local contrastive loss and introduces two augmentation strategies: random edge removal and feature masking.

The adaptive version GCA refines augmentation by assigning higher removal/masking probabilities to less important edges and features, using node centrality measures (degree, eigenvector, PageRank) to estimate importance.

Extensive experiments on datasets such as Wiki‑CS, Amazon‑Computers, Amazon‑Photo, Coauthor‑CS, and Coauthor‑Physics show that GRACE and GCA consistently outperform baseline network‑embedding methods, unsupervised GNNs, and supervised GNNs, narrowing the gap between unsupervised and supervised performance. Ablation studies confirm the benefit of adaptive augmentation over uniform strategies.

The authors conclude that local‑local contrastive objectives and importance‑aware augmentations are effective for graph self‑supervised learning, but note that the field is still early, with open questions about optimal contrastive objectives, augmentation design, and theoretical understanding.