Research map / Learning theory and algorithms / Fairness
Training individually fair ML models with sensitive subspace robustness
Mikhail Yurochkin, Amanda Bower, Yuekai Sun
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 |
|---|---|---|---|
| train with sensitive-subspace distributionally robust optimization instead of empirical risk minimization | one-layer neural network · Sentiment prediction | Race gap: 7.01$\pm$.44 -> 0.30$\pm$.05 | Improved |
| train with sensitive-subspace distributionally robust optimization instead of empirical risk minimization | one-layer neural network · Sentiment prediction | Acc.,%: 95$\pm$1 -> 94$\pm$1 | Hurt |
| use expert-provided evaluation names to define the sensitive subspace | SenSR · Sentiment prediction | Race gap: 0.30$\pm$.05 -> 0.11$\pm$.04 | Improved |
| use expert-provided evaluation names to define the sensitive subspace | SenSR · Sentiment prediction | Cuis. gap: 0.23$\pm$.05 -> 1.11$\pm$.15 | Hurt |
| project inputs onto the orthogonal complement of the sensitive subspace before training | one-layer neural network · Adult test | S-Con.: .848 -> .868 | Improved |
| train with SenSR rather than the baseline | 100-unit single-hidden-layer neural network · Adult test | S-Con.: .848 -> .934 | Improved |
| train with SenSR rather than the baseline | 100-unit single-hidden-layer neural network · Adult test | B-Acc,%: 82.9 -> 78.9 | Hurt |
| train with SenSR rather than adversarial debiasing | 100-unit single-hidden-layer neural network · Adult test | S-Con.: .807 -> .934 | Improved |
| train with adversarial debiasing rather than the baseline | 100-unit single-hidden-layer neural network · Adult test | S-Con.: .848 -> .807 | Hurt |
| use expert-provided sensitive directions rather than directions learned from a larger names dataset | SenSR · Sentiment prediction | Cuis. gap: 0.23$\pm$.05 -> 1.11$\pm$.15 | Hurt |
Closest papers by meaning
- SenSeI: Sensitive Set Invariance for Enforcing Individual Fairness (ICLR 2021 · similarity 0.92)
- Ensuring Fairness Beyond the Training Data (NeurIPS 2020 · similarity 0.89)
- Fairness via Independence: A General Regularization Framework for Machine Learning (ICLR 2026 · similarity 0.89)
- Optimal Fair Learning Robust to Adversarial Distribution Shift (ICML 2025 · similarity 0.87)
- Can Private Machine Learning Be Fair? (AAAI 2025 · similarity 0.87)
- Fairness without Harm: An Influence-Guided Active Sampling Approach (NeurIPS 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.
