DataTerminology.comNEW
Also known as: PPV Parity
A fairness criterion requiring equal positive predictive value across groups.
DESCRIPTION
Predictive parity requires that a model’s positive predictive value be equal for different demographic groups. It focuses on calibration of positive predictions across subpopulations. Achieving predictive parity may conflict with other fairness criteria when base rates differ. It is often assessed using confusion-matrix-derived metrics and group-wise evaluation. Governance teams use it to document trade-offs and justify fairness choices.
ALTERNATIVE DEFINITIONS
Wikipedia
Predictive parity requires equal positive predictive value across groups.
Concise definition situating predictive parity among multiple fairness criteria.
ArXiv (Chouldechova)
We show that predictive parity generally cannot be satisfied simultaneously with error rate balance when base rates differ.
Highlights incompatibility of fairness metrics under differing base rates.
Fairlearn
Fairness assessment compares metrics across groups, including measures such as calibration and predictive value.
Emphasizes group-wise metric comparison including calibration-related measures.
RELATED TERMS
Calibration Precision Fairness Metrics Bias Mitigation Confusion Matrix Group Fairness Model Evaluation Base Rate
COMPETING TERMS
Equalized Odds Demographic Parity
KNOWLEDGE AREAS
AI Governance Data Science AI Testing & Validation
TARGET AUDIENCE
Chief AI Officer Data Scientist AI Engineer Data Governance Manager MLOps Engineer Chief Data Officer
EXTERNAL SOURCES
- Wikipedia - Fairness (machine learning)
- IBM AIF360 - Fairness metrics documentation
- Fairlearn - Assessing fairness across groups
- ArXiv (Chouldechova) - Trade-offs among fairness metrics
TOP QUESTIONS
What does Predictive Parity measure in models?
Predictive Parity measures whether the positive predictive value is equal across demographic groups, indicating similar correctness of positive predictions for each group.
Category: Definition
Why can Predictive Parity conflict with Equalized Odds?
Differences in base rates can create incompatibilities, making simultaneous satisfaction of predictive parity and error rate balance mathematically impossible in many cases.
Category: Misunderstanding
How do you calculate Predictive Parity?
Compute positive predictive value (precision for positives) for each group, then compare for parity using thresholds or disparity ratios.
Category: Metrics
When should Predictive Parity be prioritized?
Prioritize it when the correctness of positive decisions across groups is most critical, such as approvals or diagnoses where false positives carry similar costs.
Category: Opinion
What are common mistakes when applying Predictive Parity?
Ignoring base rate differences, small sample sizes, or label noise can yield misleading parity conclusions and unintended harms.
Category: Mistakes
How is Predictive Parity documented in AI Governance?
Include group-wise PPV metrics, rationale for chosen thresholds, trade-off analysis versus other criteria, and mitigation actions in model cards or risk assessments.
Category: Management
What hot topics relate to Predictive Parity today?
Debates on multi-metric optimization, synthetic data for fairness, and post-processing calibration to approach parity without retraining are prominent.
Category: Topic
What risks come from focusing only on Predictive Parity?
Satisfying parity may worsen false negative disparities or overall accuracy, producing net harm to certain groups.
Category: Risks
How can teams monitor Predictive Parity in production?
Track PPV by group over time, set alerts for parity thresholds, and periodically recalibrate models with fresh data.
Category: Observability
What is an alternative metric to Predictive Parity?
Equalized Odds focuses on equal false positive and false negative rates across groups, often competing with calibration-oriented criteria.
Category: Alternatives