Fairness Metric

Also known as statistical parity, this metric requires the probability of a positive prediction to be identical across all groups.

• Constraint: P(ŷ=1 | A=a) = P(ŷ=1 | A=b)
• Use Case: Loan approval where access to opportunity should be equalized.
• Weakness: Ignores the ground truth. It can force the model to approve unqualified applicants in one group to match the approval rate of a more qualified group, a phenomenon known as levelling down.

This metric requires the model to have equal True Positive Rates and False Positive Rates across groups, conditioning on the actual outcome.

• Constraint: P(ŷ=1 | Y=1, A=a) = P(ŷ=1 | Y=1, A=b) AND P(ŷ=1 | Y=0, A=a) = P(ŷ=1 | Y=0, A=b)
• Use Case: Criminal recidivism prediction where both false alarms and missed risks must be balanced.
• Advantage: Matches error rates to actual outcomes, avoiding the levelling down problem of demographic parity.

A relaxation of equalized odds, this metric focuses solely on equalizing the True Positive Rate across groups.

• Constraint: P(ŷ=1 | Y=1, A=a) = P(ŷ=1 | Y=1, A=b)
• Use Case: College admissions where the goal is to ensure equally qualified candidates from different groups have the same chance of admission.
• Distinction: It permits differences in false positive rates, meaning it only guarantees fairness for the 'advantaged' outcome class.

Also called test-fairness, this metric requires the model's positive predictive value (precision) to be equal across groups.

• Constraint: P(Y=1 | ŷ=1, A=a) = P(Y=1 | ŷ=1, A=b)
• Use Case: Targeted marketing where the cost of a false positive is high and the business needs to ensure a predicted conversion is equally reliable regardless of demographic.
• Trade-off: It is mathematically impossible to satisfy both predictive parity and equalized odds simultaneously unless the base rates of the outcome are identical across groups.

A philosophical departure from group metrics, this principle mandates that similar individuals should receive similar predictions.

• Formalization: D( f(x_i), f(x_j) ) ≤ d(x_i, x_j) for a task-specific distance metric d.
• Challenge: Requires defining a fairness metric that accurately captures task-relevant similarity, which is often a complex ontological problem.
• Contrast: Addresses the fairness gerrymandering problem where group metrics are satisfied by giving positive outcomes to one subset of a group and negative to another.

A legal metric derived from the 80% rule in US employment law, measuring the ratio of favorable outcomes for a protected group relative to a reference group.

• Calculation: (Selection Rate for Protected Group) / (Selection Rate for Reference Group)
• Threshold: A ratio below 0.8 indicates potential adverse impact.
• Application: Used in hiring algorithms to flag legally actionable discrimination before deployment.

The 80% rule or four-fifths rule is a legal benchmark originating from U.S. employment law. It calculates the ratio of the favorable outcome rate for a protected group to that of a reference group.

• Formula: (Selection Rate for Protected Group) / (Selection Rate for Reference Group)
• Threshold: A ratio below 0.80 signals potential adverse impact
• Use Case: Often the first statistical test in a fairness audit before applying more nuanced metrics like equalized odds

A stricter fairness criterion requiring that a model's true positive rate and false positive rate are equal across all protected groups. Unlike demographic parity, it conditions on the ground truth label.

• Constraint: The predictor must be independent of the sensitive attribute, given the actual outcome
• Trade-off: Satisfying equalized odds often reduces overall model accuracy, forcing an explicit calibration of the fairness-utility frontier
• Origin: Formalized by Hardt, Price, and Srebro in their 2016 NeurIPS paper

Also called statistical parity, this metric requires that the probability of receiving a positive prediction is identical across all demographic groups. It is a purely outcome-based measure that ignores ground truth labels.

• Strength: Simple to explain to regulators and does not require labeled truth data for the protected attribute
• Weakness: Can lead to levelling down—reducing outcomes for the advantaged group rather than improving them for the disadvantaged
• Legal Context: Aligns with disparate impact theory under Title VII of the Civil Rights Act

A causal fairness technique that answers: What minimal change to this individual's features would flip the model's decision? It exposes whether a protected attribute is the decisive factor.

• Mechanism: Generates the closest counterfactual data point that receives a different prediction
• Fairness Application: If changing only the sensitive attribute (e.g., race or gender) alters the outcome, the model exhibits individual-level discrimination
• Tooling: Implemented in libraries like DiCE (Diverse Counterfactual Explanations) and Alibi

The broader discipline encompassing the identification and mitigation of statistical bias. Fairness metrics are the quantitative instruments within this practice.

• Detection Phase: Applying metrics like demographic parity, equalized odds, and predictive parity to sliced evaluation data
• Mitigation Strategies:
  • Pre-processing: Reweighting or resampling training data
  • In-processing: Adding fairness constraints to the loss function
  • Post-processing: Calibrating decision thresholds per group
• Governance: Requires integration with model cards and algorithmic impact assessments for full regulatory compliance

A fairness constraint mandating that the model's overall prediction accuracy is equal across demographic groups. It ensures no group systematically experiences higher error rates.

• Limitation: Accuracy parity can be satisfied even when the model exhibits significant differences in false positive and false negative rates between groups
• Complementary Metric: Must be used alongside equalized odds or predictive parity to catch masked disparities
• Example: A model with 90% accuracy for both groups might still deny loans to qualified minority applicants at twice the rate of majority applicants