Fairness drift monitoring: design framework

Ситуация: ML model launched 6 months ago. Fairness audit found bias growing. Need continuous monitoring — not periodic.

Hiring scoring model. Sensitive attributes: gender, age. Compliance team requires.

Доступные данные

• predictions: id, ts, score, applicant_id
• applicants: id, gender, age, outcome
• fairness_metrics: ts, metric, value

Задачи

• Metric selection: demographic parity, equalized odds, calibration.
• Monitoring cadence: real-time vs daily.
• Alert thresholds.
• Action playbook on alert.

Все кейсы для подготовки → ← Все кейсы