Research map / Learning theory and algorithms / Fairness
A Causal Lens for Learning Long-term Fair Policies
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 causal decomposition of long-term fairness in RL (qualification gain parity) into direct, indirect, and spurious policy effects, links the direct effect to benefit fairness, and balances both via a PPO-based constrained objective.
Problem
Long-term fairness in sequential decision-making: policies can create group disparities in cumulative qualification gains over time, and these may conflict with short-term fairness notions such as benefit fairness.
Approach
Defines qualification gain parity as the difference in expected cumulative qualification gain across groups, decomposes it via path-specific hypothetical policies into direct (DPE), indirect (IPE), and spurious (SPE) policy effects, shows DPE relates to benefit fairness, and adds a benefit-based Gini-style penalty to PPO.
Key results
- PPO-C is fairly competitive in utility with baselines PPO, A-PPO and F-PPO-L while reducing qualification gain disparity more than the baselines.
- PPO-Cb exhibits a more pronounced reduction in DPE than IPE, and increasing beta_Lambda improves benefit fairness and lowers DPE.
- Both PPO-C and PPO-Cb yield smaller Wasserstein distances between group qualification distributions than baselines when baseline differences are considered.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use PPO-C to optimize qualification gain disparity | PPO policy optimization · Settings 1, 2, and 3 | qualification gain disparity | Helped |
| use PPO-Cb to optimize qualification gain disparity | PPO policy optimization · Settings 1, 2, and 3 | qualification gain disparity | Helped |
| add the qualification gain parity constraint | PPO · Settings 1, 2, and 3 | utility | No clear effect |
| add benefit fairness regularization | PPO-Cb · Settings 1, 2, and 3 | utility | Hurt |
| add benefit fairness regularization | PPO-C · Setting 1 | qualification gain disparity | Helped |
| add benefit fairness regularization | PPO-C · Settings 2 and 3 | qualification gain disparity | No clear effect |
| increase the benefit fairness coefficient βΛ (ablation) | PPO-Cb · Settings 1 and 2 | benefit fairness | Helped |
| increase the benefit fairness coefficient βΛ (ablation) | PPO-Cb · Settings 1 and 2 | DPE | Helped |
| add benefit fairness regularization | PPO-C · Setting 1 | DPE | Helped |
| add benefit fairness regularization | PPO-C · Settings 1 and 2 | IPE | No clear effect |
| use PPO-C | PPO · Settings 1 and 2 | Wasserstein distance between qualification distributions | Helped |
| use PPO-Cb | PPO · Settings 1 and 2 | Wasserstein distance between qualification distributions | Helped |
Surprising findings
- PPO-Cb outperforms PPO-C in Setting 1 in qualification gain parity, while the two perform similarly in Settings 2 and 3.
- Benefit fairness may align with long-term fairness under certain conditions, as the loan rate becomes more balanced with larger beta_Lambda.
Limitations
- Experiments use only a simulated loan environment with three settings.
- Qualification gain function design is treated as given and its impact on benefit differences is left for future work.
- Estimating the benefit and counterfactual qualifications requires modeling assumptions such as no confounding between S and X.
Experimental setup
Datasets: Home Credit Default Risk Lending Club
Benchmarks: simulated loan environment Settings 1-3
Metrics: utility (bank profit) qualification gain disparity benefit fairness Wasserstein distance between qualification distributions loan rate balance
Baselines: PPO A-PPO F-PPO-L
Builds on: PPO benefit fairness path-specific effects performative prediction
Closest papers by meaning
- Adapting Static Fairness to Sequential Decision-Making: Bias Mitigation Strategies towards Equal Long-term Benefit Rate (ICML 2024 · similarity 0.91)
- Remembering to Be Fair: Non-Markovian Fairness in Sequential Decision Making (ICML 2024 · similarity 0.90)
- Fairness-Accuracy Trade-Offs: A Causal Perspective (AAAI 2025 · similarity 0.88)
- FairPFN: A Tabular Foundation Model for Causal Fairness (ICML 2025 · similarity 0.88)
- Fair Off-Policy Learning from Observational Data (ICML 2024 · similarity 0.87)
- Fair Participation via Sequential Policies (AAAI 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.