Research map / Probabilistic and scientific modeling / Causal discovery
Interventional Fairness on Partially Known Causal Graphs: A Constrained Optimization Approach
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 ε-IFair, a constrained-optimization method that achieves interventional fairness when the causal graph is only partially known (an MPDAG), by modeling the predictor as a child of all observed variables and using MMD over Monte Carlo interventional data.
Problem
Achieving interventional fairness when the true causal DAG is unknown and only a partially directed graph (MPDAG) learned from observational data plus background knowledge is available.
Approach
Model the predictor as a function of all observed variables, which augments the MPDAG with a prediction node whose edges from all variables remain an MPDAG; derive identification formulas via partial causal ordering and penalize MMD between interventional prediction distributions in a constrained objective, rather than only using definite non-descendants.
Key results
- On synthetic data, ε-IFair with suitable λ achieves low unfairness comparable to IFair while RMSE comparable to the Full model.
- IFair achieves near-zero unfairness but at significantly reduced prediction performance on synthetic graphs.
- On the UCI Student and Credit Risk datasets, ε-IFair trades off unfairness and RMSE/accuracy via λ, dominating the IFair baseline in the trade-off.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use all attributes with the constrained fairness-accuracy optimization | ε-IFair · Synthetic data | RMSE | Mixed |
| Use all attributes with the constrained fairness-accuracy optimization | ε-IFair · Synthetic data | MMD | Helped |
| Use all attributes with the constrained fairness-accuracy optimization | ε-IFair · UCI Student | MMD | Helped |
| Use all attributes with the constrained fairness-accuracy optimization | ε-IFair · UCI Student | RMSE | Mixed |
| Use all attributes with the constrained fairness-accuracy optimization | ε-IFair · Credit Risk | accuracy | Mixed |
| Predict using only definite non-descendants of the sensitive attribute | IFair · Synthetic data | MMD | Helped |
| Predict using only definite non-descendants of the sensitive attribute | IFair · Synthetic data | RMSE | Hurt |
| Predict using all attributes, including the sensitive attribute | Full · Synthetic data | RMSE | Helped |
| Predict using all attributes, including the sensitive attribute | Full · Synthetic data | MMD | Hurt |
| Predict using all attributes except the sensitive attribute | Unaware · Synthetic data | RMSE | Helped |
| Predict using all attributes except the sensitive attribute | Unaware · Synthetic data | MMD | Hurt |
| Predict using all attributes, including the sensitive attribute | Full · UCI Student | MMD | Hurt |
Limitations
- Assumes no selection bias or latent confounders.
- Counterfactual and path-specific fairness not addressed due to nested counterfactual identification challenges over MPDAGs.
- Unidentifiable cases require averaging over multiple MPDAG orientations.
Experimental setup
Datasets: UCI Student Performance Credit Risk Dataset Synthetic linear SEM data on ER DAGs
Benchmarks: UCI Student Performance Credit Risk Dataset Synthetic ER DAGs
Metrics: RMSE accuracy MMD-based interventional unfairness
Baselines: Full Unaware IFair
Builds on: interventional fairness (Chiappa 2019) Meek's rules MPDAG background knowledge (Perković et al.) partial causal ordering
Closest papers by meaning
- A Local Method for Satisfying Interventional Fairness with Partially Known Causal Graphs (NeurIPS 2024 · similarity 0.95)
- FairPFN: A Tabular Foundation Model for Causal Fairness (ICML 2025 · similarity 0.91)
- Causal Fairness under Unobserved Confounding: A Neural Sensitivity Framework (ICLR 2024 · similarity 0.89)
- Fairness-Accuracy Trade-Offs: A Causal Perspective (AAAI 2025 · similarity 0.89)
- Causal Bias Detection in Generative Artifical Intelligence (NeurIPS 2026 · similarity 0.89)
- Reconciling Predictive and Statistical Parity: A Causal Approach (AAAI 2024 · similarity 0.89)
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.