Research map / Learning theory and algorithms / Fairness
Approximate Heavily-Constrained Learning with Lagrange Multiplier Models
Harikrishna Narasimhan, Andrew Cotter, Yichen Zhou, Serena Wang, Wenshuo Guo
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| increase multiplier model capacity from common multiplier to 3 hidden layers (ablation) | Lagrange multiplier model · Communities & Crime, test | 95th percentile violation: 0.11 ± 0.015 -> 0.05 ± 0.017 | Improved |
| increase multiplier capacity from common multiplier to 3 hidden layers (ablation) | Lagrange multiplier model · Communities & Crime, test | error rate: 0.17 ± 0.006 -> 0.21 ± 0.009 | Hurt |
| add constraints (3-hidden layer multiplier vs unconstrained) | linear classifier · Communities & Crime, test | 95th percentile violation: 0.09 ± 0.012 -> 0.05 ± 0.017 | Improved |
| add constraints (3-hidden layer multiplier vs unconstrained) | linear classifier · Communities & Crime, test | error rate: 0.15 ± 0.003 -> 0.21 ± 0.009 | Hurt |
| use multiplier model instead of brute-force oracle FairFictitiousPlay | 3-hidden-layer multiplier · Communities & Crime, test | 95th percentile violation: 0.05 ± 0.017 vs 0.04 ± 0.008 | No clear effect |
| use linear multiplier instead of common multiplier (ablation) | Lagrange multiplier model · Communities & Crime, test | 95th percentile violation: 0.11 ± 0.015 -> 0.14 ± 0.004 | Hurt |
| proposed method vs SoftAssign at high noise 0.5 | fairness with noisy groups · Adult, noise 0.5 | test error rate: 0.218 ± 0.002 -> 0.189 ± 0.006 | Improved |
| proposed method vs SoftAssign at low noise 0.1 | fairness with noisy groups · Adult, noise 0.1 | equal opportunity violation: -0.048 ± 0.002 -> 0.015 ± 0.035 | Hurt |
| proposed method vs DRO at noise 0.5 | fairness with noisy groups · Adult, noise 0.5 | test error rate: 0.219 ± 0.012 -> 0.189 ± 0.006 | Improved |
| proposed method vs DRO at noise 0.3 | fairness with noisy groups · Adult, noise 0.3 | test error rate: 0.216 ± 0.010 -> 0.165 ± 0.003 | Improved |
| per-query constraints with multiplier model vs average constraint of Narasimhan et al. | ranking model · MSLR-WEB10K, test | 90th percentile violation: 0.246 ± 0.025 -> 0.195 ± 0.023 | Improved |
| per-query constraints with multiplier model vs Narasimhan et al. | ranking model · MSLR-WEB10K, test | pairwise error: 0.400 ± 0.014 -> 0.412 ± 0.011 | Hurt |
Closest papers by meaning
- Learning with Statistical Equality Constraints (NeurIPS 2025 · similarity 0.86)
- Fairness Overfitting in Machine Learning: An Information-Theoretic Perspective (ICML 2025 · similarity 0.86)
- Near-Optimal Solutions of Constrained Learning Problems (ICLR 2024 · similarity 0.86)
- fairret: a Framework for Differentiable Fairness Regularization Terms (ICLR 2024 · similarity 0.86)
- Improved Algorithms for Fair Matroid Submodular Maximization (NeurIPS 2025 · similarity 0.85)
- Intersectional Fairness in Reinforcement Learning with Large State and Constraint Spaces (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.
