Research map / Graph learning / GNN architectures
Debiasing Graph Neural Networks via Learning Disentangled Causal Substructure
Shaohua Fan, Xiao Wang, Yanhu Mo, Chuan Shi, Jian Tang
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Apply DisC to GCN instead of vanilla GCN | GCN · CMNIST-75sp unbiased testing, bias degrees 0.8/0.9/0.95 | accuracy: 50.43_{\pm 4.13} -> 82.60_{\pm 0.93}; 28.97_{\pm 4.4} -> 78.14_{\pm 2.14}; 13.50_{\pm 1.38} -> 63.47_{\pm 5.65} | Improved |
| Apply DisC to GIN instead of vanilla GIN | GIN · CMNIST-75sp unbiased testing, bias degrees 0.8/0.9/0.95 | accuracy: 57.75_{\pm 0.78} -> 82.10_{\pm 1.50}; 36.78_{\pm 5.55} -> 74.90_{\pm 1.81}; 16.04_{\pm 1.14} -> 58.58_{\pm 4.24} | Improved |
| Apply DisC to GCNII instead of vanilla GCNII | GCNII · CMNIST-75sp unbiased testing, bias degrees 0.8/0.9/0.95 | accuracy: 69.70_{\pm 1.73} -> 79.50_{\pm 2.48}; 57.68_{\pm 1.68} -> 76.00_{\pm 1.90}; 41.00_{\pm 3.75} -> 60.54_{\pm 5.33} | Improved |
| Apply DisC to GCN instead of vanilla GCN | GCN · CFashion-75sp unbiased testing | accuracy: 63.60_{\pm 0.53} -> 66.85_{\pm 1.11}; 57.22_{\pm 0.93} -> 65.33_{\pm 4.70}; 47.69_{\pm 0.42} -> 63.93_{\pm 1.50} | Improved |
| Apply DisC to GIN instead of vanilla GIN | GIN · CFashion-75sp unbiased testing | accuracy: 64.25_{\pm 0.46} -> 67.10_{\pm 1.07}; 58.03_{\pm 0.40} -> 59.90_{\pm 1.31}; 49.74_{\pm 0.60} -> 55.80_{\pm 0.36} | Improved |
| Apply DisC to GCNII instead of vanilla GCNII | GCNII · CFashion-75sp unbiased testing | accuracy: 66.68_{\pm 0.59} -> 66.47_{\pm 1.77}; 60.58_{\pm 0.28} -> 65.48_{\pm 0.70}; 53.18_{\pm 0.08} -> 61.75_{\pm 0.27} | Mixed |
| Apply DisC to GCN instead of vanilla GCN | GCN · CKuzushiji-75sp unbiased testing | accuracy: 38.45_{\pm 1.1} -> 55.53_{\pm 2.29}; 28.35_{\pm 0.79} -> 48.13_{\pm 2.59}; 20.70_{\pm 0.88} -> 36.63_{\pm 1.73} | Improved |
| Apply DisC to GIN instead of vanilla GIN | GIN · CKuzushiji-75sp unbiased testing | accuracy: 41.83_{\pm 0.78} -> 55.18_{\pm 1.00}; 30.09_{\pm 0.87} -> 41.75_{\pm 0.81}; 21.18_{\pm 1.63} -> 30.25_{\pm 1.63} | Improved |
| Apply DisC to GCNII instead of vanilla GCNII | GCNII · CKuzushiji-75sp unbiased testing | accuracy: 48.53_{\pm 0.25} -> 54.90_{\pm 1.30}; 36.23_{\pm 0.20} -> 44.73_{\pm 1.55}; 25.60_{\pm 0.76} -> 36.95_{\pm 0.70} | Improved |
| Add counterfactual sample generation to DisC without generation (ablation) | DisC · Three datasets, averaged over bias degrees | accuracy | Improved |
| Remove the learned edge mask module, using LDD-style disentanglement (ablation) | DisC · Three datasets, most cases | accuracy | Hurt |
| Use near-CE bias loss with q=0.1 rather than the tested larger q values (ablation) | DisC_GCN · CMNIST-75sp hyperparameter experiment | accuracy | Hurt |
Closest papers by meaning
- Disentangling, Amplifying, and Debiasing: Learning Disentangled Representations for Fair Graph Neural Networks (AAAI 2025 · similarity 0.90)
- Rethinking Causal Relationships Learning in Graph Neural Networks (AAAI 2024 · similarity 0.90)
- A Recipe for Causal Graph Regression: Confounding Effects Revisited (ICML 2025 · similarity 0.89)
- Out-of-Distribution Generalization on Graphs via Progressive Inference (AAAI 2025 · similarity 0.88)
- Disentangled Graph Self-supervised Learning for Out-of-Distribution Generalization (ICML 2024 · similarity 0.88)
- Theoretical and Empirical Insights into the Origins of Degree Bias in Graph Neural Networks (NeurIPS 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.
