Research map / Learning theory and algorithms / Fairness
Fair Participation via Sequential 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 Constrained Projected Gradient (CPG), a sequential method that enforces a fairness constraint on group participation-rate variance under performative (policy-induced) distribution shift, and proves convergence to a fair local optimum using first-order information.
Problem
Long-term algorithmic fairness in sequential settings where the policy itself shifts the population distribution (performative prediction), so myopic fairness constraints and gradient methods fail to reduce disparity or loss.
Approach
Treats the fairness constraint as a function of the performative distribution and uses a Fletcher-style exact penalty / Lagrangian dual estimate to build a sequential QP that includes a linearized fairness constraint, so the policy makes progress toward fairness and then decreases loss, with asymptotic convergence to a feasible local optimum.
Key results
- CPG achieved a feasible local optimum of the objective in all experiments, while RGD and MPG did so only rarely.
- On the Income task, MPG got stuck in an unfair local minimum and RGD diverged to the highest disparity, while CPG converged to the unique solution.
- On MovieLens, only CPG satisfied the fairness constraint, finding a solution at the boundary of the feasible set.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use CPG instead of RGD | Sequential fair participation policy · Income | disparity | Helped |
| Use CPG instead of MPG | Sequential fair participation policy · Income | fairness feasibility | Helped |
| Use CPG instead of RGD | Sequential fair participation policy · MovieLens | fairness feasibility | Helped |
| Use CPG instead of MPG | Sequential fair participation policy · MovieLens | fairness feasibility | Helped |
| Use CPG with a large initial step size | Sequential fair participation policy · IncomeThree | disparity | Hurt |
Surprising findings
- Myopic gradient methods (RGD) can actively increase both loss and disparity even near fair local minima.
- In IncomeThree, CPG temporarily increased disparity at intermediate timesteps due to a large initial step size.
Limitations
- Asymptotic convergence does not guarantee fairness at every time step; violations may persist for many steps.
- Gradients must be estimated from noisy samples (e.g., A/B testing), which can cause local increases in disparity and loss.
- Experiments are semi-synthetic with a modeled participation-rate function and deterministic algorithms, with no robustness study over hyperparameters.
Experimental setup
Datasets: ACS PUMS (Income, IncomeThree) MovieLens
Benchmarks: ACS PUMS Income task MovieLens mystery vs adventure task ACS PUMS IncomeThree task
Metrics: total loss participation rate variance (disparity) constraint feasibility
Baselines: Repeated Gradient Descent (RGD) Myopic Projected Gradient (MPG)
Builds on: performative prediction Fletcher's smooth exact penalty function repeated gradient descent
Closest papers by meaning
- Remembering to Be Fair: Non-Markovian Fairness in Sequential Decision Making (ICML 2024 · similarity 0.89)
- Adapting Static Fairness to Sequential Decision-Making: Bias Mitigation Strategies towards Equal Long-term Benefit Rate (ICML 2024 · similarity 0.88)
- A Causal Lens for Learning Long-term Fair Policies (ICLR 2025 · similarity 0.87)
- Policy Aggregation (NeurIPS 2024 · similarity 0.86)
- A Comprehensive View of Fairness through Distributional Stability (NeurIPS 2026 · similarity 0.86)
- FairDICE: Fairness-Driven Offline Multi-Objective Reinforcement Learning (NeurIPS 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.