Research map / Learning theory and algorithms / Fairness
Fair Off-Policy Learning from Observational Data
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 FairPol, a neural framework for fair off-policy learning from observational data that enforces action fairness via adversarial fair representation learning and value fairness (envy-free or max-min) via modified objectives, with generalization bounds.
Problem
Learning decision policies from observational data collected under a potentially discriminatory behavioral policy such that decisions are fair, both in terms of the action taken and the policy value across sensitive groups.
Approach
Defines action fairness (policy independent of sensitive attribute) and value fairness (envy-free and max-min over conditional policy values) for off-policy learning; enforces action fairness by adversarial fair representation learning (counterfactual domain confusion) and then optimizes value-fair objectives over the fair representation with DM/IPW/DR policy scores.
Key results
- On simulated data, FairPol with action fairness reduces action-unfairness from 2.42 (optimal unrestricted) to about 0.21-0.24 while keeping policy value at about 1.01-1.02 vs 1.24 for unrestricted.
- On simulated data, FairPol with max-min fairness achieves equal group values of 0.73 with action unfairness 0.00, versus the unrestricted policy's group values of 0.74 and 1.46.
- On Oregon real-world data, FairPol with action fairness lowers Spearman correlation with gender to 0.015 from 0.129 for the unrestricted policy at a small cost in overall policy value (0.130 vs 0.137).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Enforce action fairness with FairPol | Off-policy learning on simulated data · Simulated data test set | action fairness: 2.42 ± 0.20 -> 0.21 ± 0.05 | Helped |
| Enforce action fairness with FairPol | Off-policy learning on simulated data · Simulated data test set | policy value (overall): 1.24 ± 0.03 -> 1.02 ± 0.02 | Hurt |
| Use envy-free fairness | FairPol on simulated data · Simulated data test set | policy value (overall): 1.02 ± 0.02 -> 0.87 ± 0.17 | Hurt |
| Use max-min fairness | FairPol on simulated data · Simulated data test set | worst-group policy value: 0.74 ± 0.03 -> 0.73 ± 0.03 | Helped |
| Use DM instead of IPW or DR | FairPol with action fairness · Simulated data test set | policy value (overall): 1.02 ± 0.02; 1.01 ± 0.03; 1.01 ± 0.03 | No clear effect |
| Enforce action fairness with FairPol | Off-policy learning on Oregon health insurance data · Oregon health insurance experiment test set | action fairness: 0.129 ± 0.007 -> 0.015 ± 0.001 | Helped |
| Enforce action fairness with FairPol | Off-policy learning on Oregon health insurance data · Oregon health insurance experiment test set | policy value (overall): 0.137 ± 0.005 -> 0.130 ± 0.004 | Hurt |
| Use max-min fairness | FairPol on Oregon health insurance data · Oregon health insurance experiment test set | policy value (overall): 0.130 ± 0.004 -> 0.131 ± 0.008 | No clear effect |
| Use max-min fairness | FairPol on Oregon health insurance data · Oregon health insurance experiment test set | worst-group policy value: 0.093 ± 0.006 -> 0.100 ± 0.011 | Helped |
| Increase envy-free parameter λ | FairPol · Simulated data | policy value (overall) | Hurt |
| Increase envy-free parameter λ | FairPol · Simulated data | difference between subgroup policy values | Helped |
| Vary regularization parameter γ | FairPol · Simulated data | performance | No clear effect |
Surprising findings
- The max-min fair policy coincides with the optimal unrestricted policy in the toy example, so max-min fairness alone is not useful without action fairness.
Limitations
- No suitable baselines exist for fair off-policy learning over general policies, so comparisons are limited to unrestricted and oracle policies.
- Experiments are limited to a single simulated dataset and one real-world medical dataset.
- Generalization bounds depend on bounded outcomes and propensities away from 0 and 1, and bounds scale with the number of sensitive attribute values.
Experimental setup
Datasets: simulated credit-lending data (n=3000) Oregon health insurance experiment data
Benchmarks: simulated credit-lending data Oregon health insurance experiment
Metrics: policy value conditional policy value by sensitive group action fairness (expected difference in policy) Spearman rank correlation with sensitive attribute
Baselines: Optimal unrestricted policy Oracle action fairness
Builds on: counterfactual domain confusion loss doubly robust policy learning fair representation learning
Closest papers by meaning
- A Causal Lens for Learning Long-term Fair Policies (ICLR 2025 · similarity 0.87)
- Causal Fairness under Unobserved Confounding: A Neural Sensitivity Framework (ICLR 2024 · similarity 0.87)
- Learning Fair Policies for Multi-Stage Selection Problems from Observational Data (AAAI 2024 · similarity 0.87)
- Remembering to Be Fair: Non-Markovian Fairness in Sequential Decision Making (ICML 2024 · similarity 0.87)
- Adapting Static Fairness to Sequential Decision-Making: Bias Mitigation Strategies towards Equal Long-term Benefit Rate (ICML 2024 · similarity 0.86)
- FairPFN: A Tabular Foundation Model for Causal Fairness (ICML 2025 · similarity 0.86)
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.