Research map / Graph learning / Graph self-supervised learning
A Twist for Graph Classification: Optimizing Causal Information Flow in Graph Neural Networks
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 ICL, an information-theoretic causal learning framework for GNN graph classification that disentangles causal and non-causal subgraph/feature representations and balances their losses with MGDA multi-objective optimization.
Problem
GNNs for graph classification rely on spurious statistical correlations with non-causal features, which hurts out-of-distribution generalization and robustness under distribution shifts.
Approach
Decompose the mutual information I(Z;Y) into causal I(C;Y) and non-causal I(S;Y|C) terms, maximize the former and minimize the latter via attention-based disentanglement, variational bound cross-entropy, a KL-to-uniform penalty, and Pareto-balanced multi-objective gradient descent (MGDA).
Key results
- GIN + ICL reaches 81.55 average on Spurious-Motif/MNIST-75SP/Graph-SST2/Molhiv versus 81.20 for GIN + CAL and 70.59 for plain GIN.
- On TUDataset, GIN + ICL averages 79.35 accuracy versus 78.35 for GIN; GAT + ICL reaches 78.98 vs 77.88 for GAT.
- On MNIST-75SP at high bias (b=0.7 and 0.9 on Spurious-Motif), ICL outperforms CAL, e.g. GCN + ICL 75.09 vs GCN + CAL 69.38 at b=0.7.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add ICL to GAT | GAT · TUDataset MUTAG | accuracy: 88.58 ± 7.54 -> 91.02 ± 7.02 | Helped |
| Add ICL to GAT | GAT · TUDataset NCI1 | accuracy: 82.11 ± 1.43 -> 83.38 ± 1.7 | Helped |
| Add ICL to GIN | GIN · TUDataset PROTEINS | accuracy: 76.21 ± 3.83 -> 75.02 ± 3.51 | Hurt |
| Add ICL to GIN | GIN · TUDataset COLLAB | accuracy: 82.08 ± 1.51 -> 82.68 ± 1.06 | Helped |
| Add ICL to GCN | GCN · TUDataset IMDB-B | accuracy: 73.89 ± 5.74 -> 74.70 ± 5.21 | Helped |
| Add ICL to GIN | GIN · TUDataset IMDB-M | accuracy: 51.53 ± 2.97 -> 52.00 ± 4.18 | Helped |
| Add ICL to GIN | GIN · Spurious-Motif b=0.5 | precision: 82.89 ± 8.53 -> 82.95 ± 8.53 | Mixed |
| Add ICL to GIN | GIN · Spurious-Motif b=0.7 | precision: 86.86 ± 9.55 -> 89.00 ± 6.65 | Helped |
| Add ICL to GIN | GIN · Spurious-Motif b=0.9 | precision: 86.56 ± 8.91 -> 86.62 ± 5.40 | Helped |
| Add ICL to GIN | GIN · MNIST-75SP | accuracy: 84.59 ± 0.33 -> 84.46 ± 0.46 | Mixed |
| Add ICL to GIN | GIN · Graph-SST2 | accuracy: 18.74 ± 2.02 -> 19.07 ± 1.57 | Helped |
| Add ICL to GIN | GIN · Molhiv | ROC-AUC: 97.19 ± 0.07 -> 97.19 ± 0.05 | No clear effect |
Surprising findings
- Plain GCN outperforms or matches several GNN variants on Molhiv and the MNIST-75SP results show large variance across trials, so gains are often within standard deviations.
Limitations
- Gains on many TUDataset settings are within large standard deviations, so statistical significance is weak.
- The Pareto-optimality claim relies on a constrained formulation that is only loosely justified; the theoretical treatment of causal assumptions is informal.
- Evaluation is limited to a few GNN backbones (GCN, GIN, GAT) and older datasets.
Experimental setup
Datasets: Spurious-Motif MNIST-75sp Graph-SST2 Molhiv MUTAG NCI1 PROTEINS COLLAB IMDB-B IMDB-M
Benchmarks: Spurious-Motif MNIST-75sp Graph-SST2 Molhiv MUTAG NCI1 PROTEINS COLLAB IMDB-B IMDB-M
Metrics: accuracy precision ROC-AUC
Baselines: DiffPool SortPool AGNN GCN + CAL GIN + CAL GAT + CAL DIR Attention Top-k Pool SAG Pool
Builds on: CAL DIR MGDA
Closest papers by meaning
- A Recipe for Causal Graph Regression: Confounding Effects Revisited (ICML 2025 · similarity 0.86)
- Rethinking Causal Relationships Learning in Graph Neural Networks (AAAI 2024 · similarity 0.84)
- Almost Surely Asymptotically Constant Graph Neural Networks (NeurIPS 2024 · similarity 0.83)
- Towards Dynamic Message Passing on Graphs (NeurIPS 2024 · similarity 0.83)
- Out-of-Distribution Generalization on Graphs via Progressive Inference (AAAI 2025 · similarity 0.83)
- Improving Graph Neural Networks by Learning Continuous Edge Directions (ICLR 2025 · similarity 0.83)
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.