Reproduce “Graph Contrastive Invariant Learning from the Causal Perspective” on OpenResearch

About this paper

Yanhu Mo, Xiao Wang, Shaohua Fan, Chuan Shi

Graph Contrastive Invariant Learning (GCIL) leverages causal principles to learn invariant graph representations by distinguishing causal from non-causal information. The method consistently outperforms existing self-supervised graph learning approaches and achieves performance competitive with semi-supervised methods on node classification benchmarks.

Key insights

• Graph data inherently contains both causal information (relevant to downstream tasks) and non-causal information (irrelevant, spurious correlations).
• Truly invariant representations in GCL require augmentations that specifically perturb non-causal factors while preserving causal ones, a mechanism often missed by random augmentation.
• Low-frequency spectral components of a graph tend to capture causal information, while high-frequency components correspond to non-causal noise, enabling targeted causal interventions through spectral manipulation.

Summary by alphaXiv.