Research map / Learning theory and algorithms / Fairness
Long-Term Fair Decision Making through Deep Generative Models
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 DeepLF, a three-phase framework that trains an RCGAN simulating a temporal causal system and then a fair decision MLP on generated interventional data via repeated gradient descent, minimizing 1-Wasserstein long-term unfairness plus loss and local direct-discrimination penalties.
Problem
Mitigating group disparity over the long term in sequential decision-making, where decisions feed back into future features, while limiting use of the sensitive attribute.
Approach
Define long-term fairness as the 1-Wasserstein distance between interventional feature distributions of demographic groups at a large time step in a temporal causal graph, and optimize it via a causally structured RCGAN that generates observational and interventional data for performative risk minimization.
Key results
- On SimLoan and Taiwan within time range [1,10], DeepLF achieves local and long-term unfairness comparable with or better than LRLF and markedly better than MLP, MLP-DP, and MLP-EO.
- T-SNE of X^10 shows the two groups are more evenly mixed under DeepLF than under MLP.
- In the second setting (T=19, evaluated on [10,19]), DeepLF again gives the best accuracy-fairness trade-off among methods.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DeepLF instead of MLP, MLP-DP, or MLP-EO | Decision models on SimLoan and Taiwan · Generated data, time range [1,10] | local unfairness | Helped |
| Use DeepLF instead of MLP, MLP-DP, or MLP-EO | Decision models on SimLoan and Taiwan · Generated data, time range [1,10] | long-term unfairness | Helped |
| Use DeepLF instead of LRLF | Decision models on SimLoan and Taiwan · Generated data, time range [1,10] | local unfairness | Helped |
| Use DeepLF instead of LRLF | Decision models on SimLoan and Taiwan · Generated data, time range [1,10] | long-term unfairness | Helped |
| Use MLP, MLP-DP, or MLP-EO instead of DeepLF | Decision models on SimLoan and Taiwan · Generated data, time range [1,10] | accuracy | Mixed |
| Use DeepLF instead of the baseline methods | Decision models on SimLoan and Taiwan · Generated data, time range [10,19] | long-term unfairness | Helped |
Surprising findings
- Traditional static fairness constraints (MLP-DP, MLP-EO) show no clear decreasing trend in long-term unfairness despite some accuracy gains.
Limitations
- LRLF requires true causal structure equations, which may not be available in practice.
- Long-term fairness may not always be achievable through updating the decision model alone; theoretical feasibility not analyzed.
- Evaluation relies on synthetic and semi-synthetic data generated by the authors' own rules.
Experimental setup
Datasets: SimLoan (synthetic, generated) Taiwan credit default dataset (semi-synthetic initialization)
Benchmarks: SimLoan Taiwan
Metrics: accuracy local unfairness (direct discrimination) long-term unfairness (1-Wasserstein/Sinkhorn distance)
Baselines: MLP MLP-DP MLP-EO LRLF
Builds on: Pearl structural causal model performative risk minimization repeated gradient descent RCGAN path-specific effects long-term fairness (LRLF)
Closest papers by meaning
- Achieving Long-Term Fairness in Sequential Decision Making (AAAI 2022 · similarity 0.89)
- Adapting Static Fairness to Sequential Decision-Making: Bias Mitigation Strategies towards Equal Long-term Benefit Rate (ICML 2024 · similarity 0.87)
- A Causal Lens for Learning Long-term Fair Policies (ICLR 2025 · similarity 0.87)
- Some Optimizers are More Equal: Understanding the Role of Optimizers in Group Fairness (NeurIPS 2025 · similarity 0.86)
- Fairness without Harm: An Influence-Guided Active Sampling Approach (NeurIPS 2024 · similarity 0.85)
- FairPFN: A Tabular Foundation Model for Causal Fairness (ICML 2025 · similarity 0.85)
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.