Research map / Learning theory and algorithms / Fairness
A Unified Approach to Fair Online Learning via Blackwell Approachability
Evgenii E Chzhen, Christophe Giraud, Gilles Stoltz
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 |
|---|---|---|---|
| Use Blackwell approachability with sequential estimates of the unknown context distribution | Fair online learning with stochastic contexts · General approachability setting | distance to target set | Improved |
| Combine vanilla calibration with demographic parity | Online calibration · Vanilla calibration with demographic parity | calibration error and demographic parity violation: 1/N | Improved |
| Require group-wise no-regret when the sensitive context is irrelevant or identifiable from the non-sensitive context | Online no-regret · Group-wise no-regret | group-wise regret | Improved |
| Require group-wise no-regret with overlapping group-conditional context supports in the counterexamples | Online no-regret · Group-wise no-regret | group-wise regret | Hurt |
| Allow a demographic-parity violation level of δτ and optimize the calibration tolerance | Group-wise calibration · Nature observes sensitive context | calibration error: 1 - τ · TV(Q⁰, Q¹) ≤ ε*(δτ) ≤ 1 - τ · TV(Q⁰, Q¹) + 1/N | Mixed |
| Allow a demographic-parity violation level of δτ and optimize the calibration tolerance | Group-wise calibration · Nature does not observe sensitive context | calibration error: (1 - τ) · TV(Q⁰, Q¹) ≤ ε*(δτ) ≤ (1 - τ) · TV(Q⁰, Q¹) + 1/N | Mixed |
| Set τ = 1 when Nature does not observe the sensitive context | Group-wise calibration · Group-wise calibration with demographic parity | calibration error: 1/N | Improved |
| Require both calibration error and demographic-parity violation to be at most 1/N when Nature observes the sensitive context | Group-wise calibration · Group-wise calibration with demographic parity | calibration error: ε = 1/N | Hurt |
Closest papers by meaning
- Online Learning with an Unknown Fairness Metric (NeurIPS 2018 · similarity 0.86)
- Honor Among Bandits: No-Regret Learning for Online Fair Division (NeurIPS 2024 · similarity 0.84)
- Monotone Individual Fairness (ICML 2024 · similarity 0.83)
- Extensive-Form Game Solving via Blackwell Approachability on Treeplexes (NeurIPS 2024 · similarity 0.83)
- Remembering to Be Fair: Non-Markovian Fairness in Sequential Decision Making (ICML 2024 · similarity 0.83)
- Improved Regret Bounds for Online Fair Division with Bandit Learning (AAAI 2025 · similarity 0.82)
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.
