Definitions

In machine learning, fairness refers to the absence of any prejudice or favoritism toward an individual or group based on their inherent or acquired characteristics; many formal criteria (e.g., demographic parity, equalized odds, equal opportunity) are used to assess it.

Wikipedia: Fairness (machine learning)

Fairness is a key characteristic of trustworthy AI systems and requires identifying, measuring, and mitigating harmful bias across the AI lifecycle, including data, models, and human processes.

NIST AI Risk Management Framework

IEEE 7003 provides processes to identify, document, measure, and mitigate algorithmic bias so that AI-enabled systems achieve context-appropriate fairness toward affected persons and groups.

IEEE 7003-2024 Standard for Algorithmic Bias Considerations

Scope and Intent

  • Aim: ensure outcomes and error rates are not unjustifiably different across protected groups.
  • Typical metrics: disparate impact (80% rule), demographic parity difference, equal opportunity (TPR parity), equalized odds (TPR/FPR parity), calibration within groups.
  • Lifecycle: apply mitigation pre-/in-/post‑processing and monitor drift and disparities in production.

See Also (relationships)

  • Bias Mitigation: the set of techniques (pre‑/in‑/post‑processing, monitoring) used to achieve fairness targets defined for a system.
  • Explainability: helps diagnose sources of unfairness (feature attributions, counterfactuals) and justify mitigation choices to stakeholders.
  • Transparency: disclosures (model cards, data sheets, per‑group metrics) that enable external review of fairness goals and outcomes.
  • Accountability: governance assigning responsibility for fairness policies, reviews, approvals, and remediation when disparities arise.

Metrics (brief)

  • Disparate Impact (80% rule): ratio of selection rates between protected and reference groups; example: 0.78 (< 0.80) flags possible adverse impact.
  • Demographic Parity Difference: difference in selection rates; example: 0.10 means a 10‑point gap in positive decisions.
  • Equal Opportunity: difference in true positive rates (TPR/recall) for the qualified class; example: TPR_A − TPR_B ≤ 0.05.
  • Equalized Odds: both TPR and FPR are similar across groups; enforce combined bounds (e.g., ≤ 0.05 each).
  • Calibration within Groups: predicted probabilities match observed frequencies per group (assess via reliability curves, ECE/Brier).
  • Predictive Parity (PPV parity): positive predictive value similar across groups; note trade‑offs with other criteria when base rates differ.