Research map / Learning theory and algorithms / Fairness
Fair Bilevel Neural Network (FairBiNN): On Balancing fairness and accuracy via Stackelberg Equilibrium
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 FairBiNN, a bilevel (Stackelberg) deep learning framework that treats accuracy as the leader and demographic-parity fairness as the follower, claiming Pareto-optimal solutions and better accuracy-fairness trade-offs than Lagrangian regularization on UCI Adult and Heritage Health.
Problem
Fairness-accuracy trade-off in classification: Lagrangian/regularization approaches to group fairness may ignore interdependencies between accuracy and fairness objectives and lose information.
Approach
Formulates fair classifier training as a bilevel optimization problem (accuracy as upper-level, fairness as lower-level with unrolled/alternating gradient updates), with theoretical claims that the Stackelberg equilibrium is Pareto optimal and that its primary loss bound is at most that of the Lagrangian approach.
Key results
- On UCI Adult, FairBiNN (red diamond Pareto points) reportedly significantly outperforms benchmark fairness methods in accuracy vs. demographic parity trade-off.
- On Heritage Health, FairBiNN closely competes with recent state-of-the-art methods (per the paper's own description).
- Theory claims the bilevel primary loss upper bound is at most the Lagrangian loss under stated assumptions (Theorem on overparameterized primary loss).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use FairBiNN bilevel optimization instead of benchmark fairness methods | FairBiNN · UCI Adult | accuracy and demographic parity difference | Helped |
| use FairBiNN bilevel optimization instead of benchmark fairness methods | FairBiNN · Heritage Health | accuracy and demographic parity difference | Mixed |
| use bilevel optimization instead of Lagrangian regularization | FairBiNN · UCI Adult and Heritage Health | BCE loss | Helped |
Limitations
- Assumptions (strict convexity near optima, Lipschitz continuity, bounded layer outputs, overparameterization) may not hold in practice; softmax and attention are not Lipschitz so multiclass and attention models are excluded.
- Theory gives only relative guarantees versus regularization, not absolute fairness guarantees for the final model.
- Only demographic parity is implemented in experiments; convergence properties are not analyzed.
Experimental setup
Datasets: UCI Adult Heritage Health
Benchmarks: UCI Adult Heritage Health
Metrics: accuracy demographic parity difference BCE loss
Baselines: CVIB MIFR FCRL MaxEnt-ARL Adversarial Forgetting Fair Consistency Regularization Robust Fairness Regularization Lagrangian regularization
Builds on: Lagrangian fairness optimization Stackelberg game theory model-agnostic meta-learning
Closest papers by meaning
- FairNet: Dynamic Fairness Correction without Performance Loss via Contrastive Conditional LoRA (NeurIPS 2025 · similarity 0.88)
- FairSIN: Achieving Fairness in Graph Neural Networks through Sensitive Information Neutralization (AAAI 2024 · similarity 0.87)
- Efficient Fairness-Performance Pareto Front Computation (NeurIPS 2025 · similarity 0.86)
- fairret: a Framework for Differentiable Fairness Regularization Terms (ICLR 2024 · similarity 0.86)
- FairTune: Optimizing Parameter Efficient Fine Tuning for Fairness in Medical Image Analysis (ICLR 2024 · similarity 0.86)
- Achievable Fairness on Your Data With Utility Guarantees (NeurIPS 2024 · 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.