Research map / Graph learning / Graph self-supervised learning
Graph Contrastive Invariant Learning from the Causal Perspective
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes GCIL, a causal-perspective graph contrastive learning method that uses spectral graph augmentation as intervention on non-causal factors, plus an invariance objective and an HSIC-based independence objective, for node classification.
Problem
Graph contrastive learning is assumed to learn invariant representations, but random augmentations mix causal and non-causal information, so the learned representations may contain non-causal content and confounded causal factors.
Approach
Analyze GCL with a structural causal model, generate one view by spectral (low-frequency-preserving) graph augmentation that intervenes on non-causal high-frequency content, and train with a dimension-wise invariance loss (match per-dimension mean and std across views) plus a decorrelation loss approximating HSIC independence between dimensions.
Key results
- GCIL reaches 83.8% Ma-F1 and 84.4% Mi-F1 on Cora, outperforming the best self-supervised baseline CCA-SSG (82.9/83.6).
- On Citeseer, GCIL achieves 69.1 Ma-F1 and 73.7 Mi-F1, exceeding CCA-SSG (67.9/73.1).
- On Flickr, GCIL is weaker than GCN on Macro-F1 (40.0 vs 48.9) but best among self-supervised methods on Micro-F1 (43.0).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add the independence objective (ablation) | GCIL · Cora | Ma-F1: 57.5 -> 83.8 | Helped |
| add the independence objective (ablation) | GCIL · Wiki-CS | Ma-F1: 34.3 -> 75.6 | Helped |
| add the invariance objective (ablation) | GCIL · Cora | Ma-F1: 79.2 -> 83.8 | Helped |
| add the invariance objective (ablation) | GCIL · Wiki-CS | Ma-F1: 57.5 -> 75.6 | Helped |
| add spectral-based augmentation (ablation) | GCIL · Cora | Ma-F1: 83.3 -> 83.8 | Helped |
| add spectral-based augmentation (ablation) | GCIL · Wiki-CS | Ma-F1: 75.1 -> 75.6 | Helped |
| increase α over 0.5 to 1.4 | GCIL · Cora and Wiki-CS | F1-Macro and F1-Micro | No clear effect |
| increase β from 0 over a suitable range | GCIL · Cora and Wiki-CS | F1-Macro and F1-Micro | Helped |
| use excessively large β | GCIL · Cora and Wiki-CS | F1-Macro and F1-Micro | Hurt |
| increase γ within tested ranges | GCIL · Cora and Wiki-CS | F1-Macro and F1-Micro | Helped |
| use excessively large γ | GCIL · Cora and Wiki-CS | F1-Macro and F1-Micro | Hurt |
Limitations
- Does not beat the semi-supervised GCN on Flickr
- Ablation and sensitivity results are reported on only Cora and Wiki-CS
- Causal assumptions (C and S separation) are asserted rather than verified
Experimental setup
Datasets: Cora Citeseer PubMed Wiki-CS Flickr
Benchmarks: Cora Citeseer PubMed Wiki-CS Flickr
Metrics: Macro-F1 Micro-F1
Baselines: GCN GAT DGI MVGRL GRACE GCA GraphCL CCA-SSG COSTA ProGCL
Builds on: GRACE GraphCL CCA-SSG GCN HSIC
Closest papers by meaning
- Rethinking Dimensional Rationale in Graph Contrastive Learning from Causal Perspective (AAAI 2024 · similarity 0.91)
- Dissecting the Failure of Invariant Learning on Graphs (NeurIPS 2024 · similarity 0.90)
- Hierarchical Topology Isomorphism Expertise Embedded Graph Contrastive Learning (AAAI 2024 · similarity 0.90)
- Perfect Alignment May be Poisonous to Graph Contrastive Learning (ICML 2024 · similarity 0.90)
- Does GCL Need a Large Number of Negative Samples? Enhancing Graph Contrastive Learning with Effective and Efficient Negative Sampling (AAAI 2025 · similarity 0.90)
- ReGCL: Rethinking Message Passing in Graph Contrastive Learning (AAAI 2024 · similarity 0.89)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.