Research map / Probabilistic and scientific modeling / Causal discovery
FairPFN: A Tabular Foundation Model for Causal Fairness
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 FairPFN, a tabular foundation model pre-trained on synthetic causal fairness data from sparse MLP-based SCMs to remove the causal effect of a binary exogenous protected attribute from predictions using only observational data.
Problem
Causal fairness methods require a known causal model, which is often unidentifiable from observational data, limiting their practical use and risking fairwashing from incorrect causal graphs.
Approach
Pre-train a PFN-style transformer on a prior of synthetic causal fairness datasets (MLP-based SCMs with a binary exogenous protected attribute), with loss computed against fair targets generated by dropping the protected attribute's outgoing edges, so at inference it outputs causally fair predictions from observational data alone, approximating a posterior predictive distribution over causal expla
Key results
- FairPFN achieves an average rank of 1.88 out of 4 on prediction ATE across 600 synthetic datasets, and is Pareto optimal on 40% of them using only observational data.
- On real-world data FairPFN achieves prediction ATE below 0.01 on both Law School and Adult while keeping accuracy comparable to Unfair.
- FairPFN achieves the lowest median and maximum counterfactual absolute error on both real-world datasets, with maximum divergence under 0.05 on Law School and 0.2 on Adult.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use FairPFN instead of baselines without causal information | Synthetic causal case studies · Synthetic causal case studies | prediction ATE | Helped |
| Use FairPFN rather than EGR | Real-world datasets · Law School Admissions and Adult Census Income | counterfactual absolute error: less than 0.05 on the Law School dataset and 0.2 on the Adult Census Income dataset | Helped |
| Increase dataset size (ablation) | FairPFN · Synthetic causal case studies | prediction ATE | Helped |
| Increase dataset noise (ablation) | FairPFN · Synthetic causal case studies | prediction ATE | No clear effect |
| Increase the protected attribute's base causal effect (ablation) | FairPFN · Synthetic causal case studies | prediction ATE | Mixed |
| Increase graph complexity (ablation) | FairPFN · Synthetic prior samples | predictive accuracy | Hurt |
| Use multiple protected attributes, outside the single-attribute prior (ablation) | FairPFN · Synthetic data with multiple protected attributes | prediction ATE | Hurt |
| Use an endogenous protected attribute, outside the exogenous-attribute prior (ablation) | FairPFN · Synthetic data with endogenous protected attributes | prediction ATE | Hurt |
Surprising findings
- The paper's Cntf. Avg. baseline (using counterfactual data) outperforms the causal-information CFP baseline in error, which the authors attribute to access to more observables and PFN use.
- FairPFN over-corrects and produces negative prediction ATE when the protected attribute's base causal effect is high.
Limitations
- Limited to a single binary exogenous protected attribute and binary classification; violating these assumptions reverts FairPFN to a normal classifier.
- Assumes protected attributes are exogenous, so it cannot separate legally admissible spurious effects from direct/indirect effects.
- Higher variance on small datasets with fewer than 250 samples.
Experimental setup
Datasets: Law School Admissions (1998 LSAC National Longitudinal Bar Passage Study) Adult Census Income synthetic causal case studies
Benchmarks: synthetic causal case studies (6 groups, 600 datasets) Law School Admissions Adult Census Income
Metrics: average treatment effect (ATE) of predictions absolute error (AE) between observational and counterfactual 1-AUC statistical parity (DSP) Pareto front
Baselines: Unfair Unaware Cntf. Avg. Constant Random CFP EGR
Builds on: TabPFN Prior-data Fitted Networks Counterfactually Fair Prediction Causal Fairness Analysis Exponentiated Gradient Reduction
Closest papers by meaning
- Fairness-Accuracy Trade-Offs: A Causal Perspective (AAAI 2025 · similarity 0.92)
- Interventional Fairness on Partially Known Causal Graphs: A Constrained Optimization Approach (ICLR 2024 · similarity 0.91)
- Causal Fairness under Unobserved Confounding: A Neural Sensitivity Framework (ICLR 2024 · similarity 0.90)
- Causal Bias Detection in Generative Artifical Intelligence (NeurIPS 2026 · similarity 0.90)
- Reconciling Predictive and Statistical Parity: A Causal Approach (AAAI 2024 · similarity 0.90)
- A Local Method for Satisfying Interventional Fairness with Partially Known Causal Graphs (NeurIPS 2024 · similarity 0.90)
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.