Research map / Graph learning / GNN architectures
Rethinking Causal Relationships Learning 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.
Builds a synthetic CRCG graph dataset with controllable causal and confounding structure, analyzes existing GNN causal-enhancement methods on it, and proposes R-CAM, a lightweight plug-and-play representation-level loss that emphasizes consistent features and suppresses confounded ones.
Problem
Existing GNN causal-enhancement methods are not well understood: it is unclear when they help versus hurt, since GNNs trained on graph data can absorb confounders that are probabilistically correlated with labels.
Approach
Analyze causal-enhancement GNNs on a controllable synthetic dataset, conclude that their gains depend on the data prior and confounder-label correlation, and therefore apply causal operations directly to output node representations via an anchor-based feature emphasis loss and a deceptive-feature ignoring loss (R-CAM), which is used only during training.
Key results
- Adding R-CAM improved DIR on Graph-SST5 from 44.20 to 46.00 and on Spurious-Motif from 43.60 to 47.30.
- On CRCG (P=20%), ERM+R-CAM reached 31.60 vs 28.80 for ERM, and ASAP+R-CAM reached 30.80 vs 26.05 for ASAP.
- In the no-confounder experiments, CIGA scored 5.15 points below ERM at noise=0 and RCGRL scored 4.66 points below ERM at noise=1, showing causal-enhancement methods can underperform ERM.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add R-CAM to ERM | ERM · Graph-SST5, Graph-Twitter, Spurious-Motif, and CRCG (P=20%, P=40%, P=80%) | accuracy: 42.30±0.87 -> 43.40±0.68; 61.20±1.05 -> 63.70±1.21; 33.20±0.95 -> 35.60±1.05; 28.80±0.75 -> 31.60±0.93; 25.94±1.63 -> 25 | Helped |
| Add R-CAM to ASAP | ASAP · Graph-SST5, Graph-Twitter, Spurious-Motif, and CRCG (P=20%, P=40%, P=80%) | accuracy: 44.50±1.34 -> 46.00±0.97; 61.50±0.97 -> 64.10±0.52; 34.90±1.25 -> 34.20±1.37; 26.05±1.26 -> 30.80±1.32; 23.62±1.23 -> 28 | Mixed |
| Add R-CAM to DIR | DIR · Graph-SST5, Graph-Twitter, Spurious-Motif, and CRCG (P=20%, P=40%, P=80%) | accuracy: 44.20±1.26 -> 46.00±1.60; 62.80±0.97 -> 62.40±0.52; 43.60±0.73 -> 47.30±1.47; 23.60±0.84 -> 31.30±1.47; 27.15±0.86 -> 30 | Mixed |
| Add R-CAM to CIGA | CIGA · Graph-SST5, Graph-Twitter, Spurious-Motif, and CRCG (P=20%, P=40%, P=80%) | accuracy: 44.20±1.03 -> 45.40±0.82; 58.90±0.77 -> 60.70±1.31; 34.40±0.79 -> 36.00±1.03; 27.40±0.98 -> 27.00±0.89; 26.06±0.86 -> 36 | Mixed |
| Add R-CAM to RCGRL | RCGRL · Graph-SST5, Graph-Twitter, Spurious-Motif, and CRCG (P=20%, P=40%, P=80%) | accuracy: 44.50±1.46 -> 46.50±1.08; 60.10±0.74 -> 63.40±0.96; 45.70±0.98 -> 48.50±0.75; 33.30±0.93 -> 34.40±1.39; 24.83±0.69 -> 28 | Helped |
| Add R-CAM to DISC | DISC · Graph-SST5, Graph-Twitter, Spurious-Motif, and CRCG (P=20%, P=40%, P=80%) | accuracy: 34.40±1.28 -> 38.15±1.21; 62.50±1.54 -> 61.70±0.89; 42.85±1.23 -> 47.52±0.97; 40.10±1.36 -> 41.05±1.05; 37.00±0.92 -> 39 | Mixed |
Surprising findings
- Causal-enhancement GNNs (e.g., CIGA, RCGRL) sometimes perform worse than conventional ERM, including on confounder-free data.
- The advantage of causal-enhancement methods over ERM shifts with the confounder-label correlation probability P rather than being consistent.
Limitations
- Gains from R-CAM are inconsistent; for example, on Graph-Twitter DIR+R-CAM is slightly below DIR and on CRCG CIGA+R-CAM is not improved at P=20%.
- The hyperparameter tau must be tuned per dataset, since the best value varies across datasets.
- The theoretical results rely on assumptions (Markov generation, conditional independence of ancestors given G) that are strong and only partially validated.
Experimental setup
Datasets: CRCG Graph-SST5 Graph-Twitter Spurious-Motif
Benchmarks: CRCG Graph-SST5 Graph-Twitter Spurious-Motif
Metrics: classification accuracy CPU time overhead Friedman test statistic
Baselines: ERM ASAP DIR CIGA RCGRL DISC
Builds on: DIR CIGA DISC ASAP Spurious-Motif
Closest papers by meaning
- A Recipe for Causal Graph Regression: Confounding Effects Revisited (ICML 2025 · similarity 0.91)
- Out-of-Distribution Generalization on Graphs via Progressive Inference (AAAI 2025 · similarity 0.89)
- LLM Enhancers for GNNs: An Analysis from the Perspective of Causal Mechanism Identification (ICML 2025 · similarity 0.89)
- Graph Contrastive Invariant Learning from the Causal Perspective (AAAI 2024 · similarity 0.88)
- Causal Concept Graph Models: Beyond Causal Opacity in Deep Learning (ICLR 2025 · similarity 0.88)
- Towards Learning and Explaining Indirect Causal Effects in Neural Networks (AAAI 2024 · 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.