Research map / Probabilistic and scientific modeling / Causal discovery
A Local Method for Satisfying Interventional Fairness with Partially Known Causal Graphs
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 min-max optimization framework that uses a local IDA-style enumeration of possible parental sets of the sensitive attribute on CPDAGs/MPDAGs to achieve interventional fairness while using all variables for prediction.
Problem
Achieving interventional fairness when the true causal DAG is unknown and only a partially directed graph (CPDAG or MPDAG with background knowledge) is available, without sacrificing much prediction accuracy or relying on strong identification assumptions.
Approach
Locally enumerate all possible parental sets of the sensitive attribute from the CPDAG/MPDAG (via IDA and Meek's rules), estimate propensities and treatment effects for each, and train the predictor with a min-max penalty bounding worst-case interventional unfairness, using all covariates rather than only non-descendants.
Key results
- On synthetic data (noise 1.5, 10 nodes), the proposed method reports RMSE 0.623 and unfairness 0.119, improving over Fair (0.747, 0.128) and FairRelax (0.653, 0.142) on both metrics across all settings.
- On OULAD, the method achieves RMSE 0.491 versus 0.499 for epsilon-IFair and 0.502 for Full, with unfairness 0.024.
- On COMPAS the method reports RMSE 0.219 and unfairness 0.179, against 0.235 and 0.190 for epsilon-IFair.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use the proposed min-max method instead of ϵ-IFair | Interventional-fairness prediction · COMPAS | RMSE: 0.235 ± 0.091 -> 0.219 ± 0.020 | Helped |
| Use the proposed min-max method instead of ϵ-IFair | Interventional-fairness prediction · COMPAS | Unfairness: 0.190 ± 0.120 -> 0.179 ± 0.011 | Helped |
| Use the proposed min-max method instead of ϵ-IFair | Interventional-fairness prediction · Adult | RMSE: 0.413 ± 0.010 -> 0.375 ± 0.019 | Helped |
| Use the proposed min-max method instead of Fair | Interventional-fairness prediction · Adult | Unfairness: 0.187 ± 0.205 -> 0.171 ± 0.021 | Helped |
| Use the proposed min-max method instead of ϵ-IFair | Interventional-fairness prediction · OULAD | RMSE: 0.499 ± 0.041 -> 0.491 ± 0.040 | Helped |
| Use the proposed min-max method instead of ϵ-IFair | Interventional-fairness prediction · OULAD | Unfairness: 0.027 ± 0.020 -> 0.024 ± 0.018 | Helped |
| Use the proposed method instead of Full | Synthetic prediction, noise 1.5, 10 nodes, 20 edges · Synthetic, noise 1.5, 10 nodes, 20 edges | RMSE: 0.576 ± 0.218 -> 0.623 ± 0.210 | Hurt |
| Use the proposed method instead of Fair | Synthetic prediction, noise 1.5, 10 nodes, 20 edges · Synthetic, noise 1.5, 10 nodes, 20 edges | Unfairness: 0.128 ± 0.200 -> 0.119 ± 0.175 | Helped |
| Use the proposed method instead of Full | Synthetic prediction, noise 1.5, 40 nodes, 80 edges · Synthetic, noise 1.5, 40 nodes, 80 edges | RMSE: 0.538 ± 0.183 -> 0.606 ± 0.178 | Hurt |
| Use the proposed method instead of Fair | Synthetic prediction, noise 1.5, 40 nodes, 80 edges · Synthetic, noise 1.5, 40 nodes, 80 edges | Unfairness: 0.038 ± 0.059 -> 0.036 ± 0.054 | Helped |
| Increase the fairness threshold C | Proposed min-max method · Synthetic experiments | RMSE | Helped |
| Increase the fairness threshold C | Proposed min-max method · Synthetic experiments | Unfairness | Hurt |
Surprising findings
- Using descendants of the sensitive attribute for prediction, constrained by the min-max fairness penalty, can improve accuracy while keeping interventional unfairness low.
- On COMPAS, epsilon-IFair attains lower unfairness (0.190) than the proposed method (0.179) is not the case; the proposed method is lower, but on Adult the proposed method's unfairness (0.171) is higher than epsilon-IFair (0.155).
Limitations
- Relies on the CPDAG from causal discovery and estimated propensities; inaccurate graphs or estimates can cause mild fairness violations.
- Does not handle hidden variables in partially known DAGs.
- Fairness is controlled only up to a pre-specified threshold C, so strict zero causal effect is not guaranteed.
Experimental setup
Datasets: synthetic linear SEM data (ER graphs, 10-40 nodes) COMPAS Adult OULAD
Benchmarks: synthetic linear SEM datasets COMPAS Adult OULAD
Metrics: RMSE unfairness (RMSE between do(A=a) and do(A=a') predictions) AUC precision and recall for adjustment sets
Baselines: Full Unaware Oracle FairRelax Fair epsilon-IFair
Builds on: IDA framework Zuo et al. 2022 FAIR / FairRelax Kilbertus et al. 2017 interventional fairness Meek's rules PC algorithm
Closest papers by meaning
- Interventional Fairness on Partially Known Causal Graphs: A Constrained Optimization Approach (ICLR 2024 · similarity 0.95)
- FairPFN: A Tabular Foundation Model for Causal Fairness (ICML 2025 · similarity 0.90)
- Causal Fairness under Unobserved Confounding: A Neural Sensitivity Framework (ICLR 2024 · similarity 0.89)
- Local Causal Discovery for Structural Evidence of Direct Discrimination (AAAI 2025 · similarity 0.88)
- The Fragility of Fairness: Causal Sensitivity Analysis for Fair Machine Learning (NeurIPS 2024 · similarity 0.87)
- Mind the Graph When Balancing Data for Fairness or Robustness (NeurIPS 2024 · similarity 0.87)
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.