Research map / Graph learning / Graph self-supervised learning
A Recipe for Causal Graph Regression: Confounding Effects Revisited
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 a causal graph regression framework that enhances the graph information bottleneck with a confounder-predictive term and replaces label-based causal intervention with a contrastive loss, evaluated on GOOD-ZINC and ReactionOOD.
Problem
Causal graph learning methods are designed for classification; their confounder-handling assumptions (confounders carry no predictive power) and label-dependent interventions do not transfer to graph-level regression under OOD shift.
Approach
Adds a confounding-subgraph predictive term I(S;Y) to the GIB objective, and replaces label-based backdoor-style intervention with an InfoNCE-style contrastive loss that aligns original graph representations with randomly mixed causal+confounding representations.
Key results
- On GOOD-ZINC Scaffold-Covariate, the method reaches 0.1046 OOD MAE vs 0.1419 for the next-best GSAT.
- On RDB7 under concept shift (Total Atom Number... first reactant scaffold concept OOD), the method achieves the lowest OOD RMSE of 15.73±0.37 among compared methods, per the text.
- Best OOD performance in 6 of 10 ReactionOOD cases, second in 2 cases, according to the authors.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use the full method instead of the strongest listed baseline | Ours on GOOD-ZINC · GOOD-ZINC Scaffold- and Size-domain OOD shifts | MAE: 0.1419±0.0043 -> 0.1046±0.0007; 0.0999±0.0029 -> 0.0518±0.0007; 0.2112±0.0033 -> 0.1484±0.0033; 0.1043±0.0001 -> 0.0580± | Helped |
| Use the full method instead of the strongest listed baseline | Ours on ReactionOOD · Cycloaddition, RDB7, and E2 & SN2 OOD shifts | RMSE: 4.63±0.05 -> 4.22±0.09; 4.36±0.15 -> 4.42±0.24; 14.81±0.05 -> 14.94±0.25; 4.97±0.07 -> 5.03±0.09 | Mixed |
| Include both enhanced GIB and causal intervention losses rather than removing both (ablation) | Ours on ZINC, Cycloaddition, E2SN2, and RDB7 · OOD datasets | MAE or RMSE | Helped |
| Add causal intervention loss CI (ablation) | Ours with GIB · OOD datasets | MAE or RMSE | Helped |
| Add enhanced GIB loss (ablation) | Ours with CI · OOD datasets, especially E2SN2 | MAE or RMSE | Helped |
| Add both enhanced GIB and CI losses rather than using only the causal subgraph readout loss (ablation) | Ours · OOD datasets | MAE or RMSE | Helped |
| Model confounding subgraph predictive power rather than ignoring it (ablation) | Ours without CI on GOOD-Motif size domain · GOOD-Motif OOD | accuracy | Helped |
| Use contrastive-learning causal intervention instead of predictions from randomly generated counterfactual graphs (ablation) | Ours without predictive modeling of confounders on GOOD-Motif size domain · GOOD-Motif OOD | accuracy | Helped |
| Vary alpha (ablation) | Ours on Cycloaddition · Cycloaddition OOD settings | RMSE | No clear effect |
| Increase beta beyond its useful range (ablation) | Ours on Cycloaddition · Scaffold-Covariate OOD | RMSE | Hurt |
| Set lambda in the suitable interval 0.2–0.4 rather than too large or too small (ablation) | Ours on Cycloaddition · Cycloaddition OOD settings | RMSE: 0.2–0.4 | Helped |
| Set lambda too large or too small (ablation) | Ours on Cycloaddition · Especially Size-Concept | RMSE | Hurt |
Surprising findings
- Non-causal baselines such as Coral sometimes get better ID RMSE than the proposed method while degrading OOD.
Limitations
- Causal methods generally face challenges on smaller datasets such as RDB7, per the authors.
- Gaussian assumptions with identity covariance and fixed unit conditional variance are simplifications.
- Method is not uniformly best; some ReactionOOD cases are won by baselines.
Experimental setup
Datasets: GOOD-ZINC Cycloaddition RDB7 E2&SN2
Benchmarks: GOOD-ZINC Cycloaddition RDB7 E2&SN2 GOOD-Motif
Metrics: MAE RMSE accuracy
Baselines: ERM IRM VREx Mixup DANN Coral CIGA DIR GSAT GroupDRO
Builds on: graph information bottleneck CAL InfoNCE GSAT
Closest papers by meaning
- Rethinking Causal Relationships Learning in Graph Neural Networks (AAAI 2024 · similarity 0.91)
- Out-of-Distribution Generalization on Graphs via Progressive Inference (AAAI 2025 · similarity 0.89)
- Graph Contrastive Invariant Learning from the Causal Perspective (AAAI 2024 · similarity 0.89)
- Quantifying Distributional Invariance in Causal Subgraph for IRM-Free Graph Generalization (NeurIPS 2025 · similarity 0.88)
- Rethinking Dimensional Rationale in Graph Contrastive Learning from Causal Perspective (AAAI 2024 · similarity 0.88)
- Graph Distance Based on Cause-Effect Estimands with Latents (NeurIPS 2026 · similarity 0.88)
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.