Research map / Graph learning / Graph self-supervised learning
Causal Invariance-aware Augmentation for Brain Graph Contrastive Learning
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 CIA-GCL, a causal invariant-subgraph extraction method combined with invariance-aware graph augmentation and contrastive learning for brain-graph disease classification under multi-site distribution shift.
Problem
Brain graph (fMRI functional connectivity) classification suffers from distribution shifts from multi-site collection and individual differences, which hurt generalization; existing augmentations may destroy discriminative local structure.
Approach
Learn a causal invariant subgraph via edge masks with Gumbel-Softmax, use it as the GCL anchor, and build positive views by combining the invariant subgraph with mixed spurious subgraphs from other-label subjects, trained with causal, contrastive and invariant losses.
Key results
- On ABIDE I (AAL), CIA-GCL reaches 73.42% accuracy and 74.34% average over six metrics, best among compared methods (CI-GNN 71.85% avg).
- On ABIDE II, CIA-GCL achieves 72.09% accuracy and 70.51% average, best average vs. GATE (69.37%).
- Under cross-site generalization, CIA-GCL averages 72.38% on ABIDE I target domains vs. 69.86% for CI-GNN; recall is weaker on several datasets, indicating conservative decisions.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use CIA-GCL instead of CI-GNN | Brain graph classification · ABIDE I | accuracy: 71.89±2.91 -> 73.42±2.75 | Helped |
| Use CIA-GCL instead of CI-GNN | Brain graph classification · ABIDE II | accuracy: 70.04±2.62 -> 72.09±2.47 | Helped |
| Use CIA-GCL instead of BrainNetCNN | Brain graph classification · ADHD200 | accuracy: 70.71±4.15 -> 75.21±2.57 | Helped |
| Use CIA-GCL instead of Contrasformer | Brain graph classification · ABIDE I (target-domain evaluation) | accuracy: 70.50±2.83 -> 72.03±0.90 | Helped |
| Use CIA-GCL instead of Contrasformer | Brain graph classification · ABIDE II (target-domain evaluation) | accuracy: 69.81±2.05 -> 71.51±2.94 | Helped |
| Use CIA-GCL instead of CI-GNN | Brain graph classification · ADHD200 (target-domain evaluation) | accuracy: 68.57±1.25 -> 71.32±1.05 | Helped |
| Add Lcau to ERM (ablation) | CIA-GCL · ABIDE I | accuracy: 71.13±4.01 -> 72.17±4.10 | Helped |
| Add Lcau to ERM (ablation) | CIA-GCL · ABIDE II | accuracy: 70.07±2.21 -> 69.87±3.00 | Hurt |
| Add Lcau to ERM (ablation) | CIA-GCL · ADHD200 | accuracy: 72.96±2.71 -> 74.02±4.17 | Helped |
| Add Lcon and the invariance-aware augmentation strategy (ablation) | CIA-GCL with ERM + Lcau · ABIDE I | accuracy: 72.17±4.10 -> 72.63±2.66 | Helped |
| Add Lcon and the invariance-aware augmentation strategy (ablation) | CIA-GCL with ERM + Lcau · ABIDE II | accuracy: 69.87±3.00 -> 71.41±2.48 | Helped |
| Add Lcon and the invariance-aware augmentation strategy (ablation) | CIA-GCL with ERM + Lcau · ADHD200 | accuracy: 74.02±4.17 -> 74.26±2.36 | Helped |
Surprising findings
- The method has relatively low recall on several datasets (e.g., ABIDE II recall 64.39%), which the authors attribute to conservative decisions.
- The ablation shows the causal loss alone barely changes performance on some datasets, with gains mostly from combining losses.
Limitations
- Environments for augmentation are defined by labels, which the authors note is a simple and rough distribution-shift proxy.
- Recall is low on several datasets, suggesting conservative predictions.
- Results on ADHD200 have high variance and the method is evaluated only on three rs-fMRI datasets.
Experimental setup
Datasets: ABIDE I ABIDE II ADHD200
Benchmarks: ABIDE I ABIDE II ADHD200
Metrics: accuracy precision recall F1-score balanced accuracy AUC
Baselines: RF SVM BrainNetCNN CIGA GATE Com-BrainTF BrainGB METAFormer Contrasformer BrainIB CI-GNN
Builds on: CIGA Gumbel-Softmax graph contrastive learning Edge-to-Edge/Edge-to-Node brain graph convolution (BrainNetCNN) invariant risk minimization
Closest papers by meaning
- Graph Contrastive Invariant Learning from the Causal Perspective (AAAI 2024 · similarity 0.88)
- Disentangling Invariant Subgraph via Variance Contrastive Estimation under Distribution Shifts (ICML 2025 · similarity 0.86)
- Quantifying Distributional Invariance in Causal Subgraph for IRM-Free Graph Generalization (NeurIPS 2025 · similarity 0.85)
- Community-Invariant Graph Contrastive Learning (ICML 2024 · similarity 0.85)
- SOLA-GCL: Subgraph-Oriented Learnable Augmentation Method for Graph Contrastive Learning (AAAI 2025 · similarity 0.85)
- Out-of-Distribution Generalization on Graphs via Progressive Inference (AAAI 2025 · similarity 0.84)
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.