Definition
Fairness parity means that a model’s decisions or outcomes are approximately equal across different demographic groups or cohorts, according to a chosen fairness metric.
In plain words:
The system treats groups similarly, with performance gaps kept within an acceptable tolerance.
Types of Parity (Common Fairness Metrics)
Depending on what we measure, “parity” can mean different things:
- Demographic Parity (Statistical Parity)
- Each group receives positive outcomes at the same rate.
- Example: Loan approval rate = 60% for Group A and 58% for Group B → parity (within 2pp).
- Equal Opportunity
- True Positive Rate (TPR) is equal across groups.
- Example: Model correctly identifies 80% of qualified applicants in both groups.
- Equalized Odds
- Both TPR and FPR are equal across groups.
- Predictive Parity (Calibration)
- Predicted probabilities correspond equally well to actual outcomes across groups.
“Fairness parity within Xpp”
- Often defined as: group-level metric difference ≤ X percentage points.
- Example (churn model accuracy):
- Group A accuracy = 85%
- Group B accuracy = 83%
- Gap = 2pp → fairness parity within 3pp.
Why It Matters
- Ensures no subgroup is unfairly advantaged or penalized by the model.
- Often required by regulators (e.g., EEOC’s 80% rule in hiring, GDPR in EU).
- Builds trust with customers & stakeholders.
Practical Implementation
- Pick fairness metric(s): accuracy, TPR, selection rate, calibration.
- Segment data: e.g., by gender, age, geography.
- Measure group gaps: difference between best vs worst group.
- Set tolerance: e.g., ≤ 3pp gap is acceptable.
- Monitor continuously: fairness metrics can drift over time.
Summary
- Fairness parity = model outcomes are similar across groups.
- Defined relative to a fairness metric (e.g., accuracy, TPR, selection rate).
- A tolerance (e.g., within 3pp) ensures differences are not material.