Algorithmic Fairness

Also known as statistical parity, this is a group fairness metric that requires a model's positive prediction rate to be equal across different protected groups (e.g., race, gender). It ensures the selection rate is independent of the sensitive attribute.

• Formula: P(Ŷ=1 | A=a) = P(Ŷ=1 | A=b) for all groups a, b.
• Use Case: Screening resumes where the proportion of candidates selected should be equal across demographic groups.
• Limitation: Can conflict with meritocracy if base rates of qualification differ between groups.

A fairness criterion requiring that the model's true positive rate (recall) is equal across protected groups. It focuses on ensuring qualified individuals from all groups have an equal chance of being correctly identified.

• Formula: P(Ŷ=1 | Y=1, A=a) = P(Ŷ=1 | Y=1, A=b).
• Key Insight: Only considers the actually qualified subset (Y=1).
• Example: In lending, an approved loan rate should be equal for creditworthy applicants across different racial groups.

A stricter fairness metric than Equal Opportunity. It requires that both true positive rates and false positive rates are equal across protected groups. The model's error rates must be independent of the sensitive attribute.

• Formula: P(Ŷ=1 | Y=y, A=a) = P(Ŷ=1 | Y=y, A=b) for y ∈ {0,1}.
• Implication: The model must be equally accurate for all groups.
• Trade-off: Often impossible to achieve simultaneously with high accuracy if base rates differ, leading to fairness-accuracy trade-offs.

Also known as outcome test. This metric requires that the precision (positive predictive value) of the model is equal across groups. It ensures that those who receive a positive prediction are equally likely to be correct, regardless of group membership.

• Formula: P(Y=1 | Ŷ=1, A=a) = P(Y=1 | Ŷ=1, A=b).
• Context: Critical in settings like criminal risk assessment, where the goal is for the predicted "high risk" group to have the same actual recidivism rate across demographics.
• Conflict: Known to be mathematically incompatible with Equalized Odds when prevalence differs between groups (except in perfect classifiers).

A causal fairness notion that evaluates fairness at the individual level. A model is counterfactually fair if its prediction for an individual is the same in the actual world and in a counterfactual world where the individual belonged to a different protected group, holding all else equal.

• Foundation: Based on structural causal models and do-calculus.
• Goal: To remove the direct and indirect discriminatory effects of the sensitive attribute via causal pathways.
• Application: Used in complex scenarios where historical biases are embedded in correlated features (e.g., using zip code as a proxy for race).

A legal and statistical doctrine originating from U.S. employment law (the 80% rule). It measures adverse, disproportionate outcomes on a protected class, regardless of the model's intent.

• Calculation: (Selection Rate for Disadvantaged Group) / (Selection Rate for Advantaged Group).
• Threshold: A ratio below 0.8 typically indicates evidence of disparate impact.
• Key Difference: Unlike metrics like Equalized Odds, it does not consider ground truth (Y). It is purely based on outcomes (Ŷ).
• Regulatory Context: A central concept in compliance with regulations like the U.S. Equal Employment Opportunity Commission guidelines.

Bias auditing is the systematic process of evaluating a dataset or machine learning model for the presence of unfair, discriminatory, or skewed representations across different demographic or contextual groups. It involves:

• Statistical testing for disparities in model performance metrics (e.g., false positive rates) across protected groups.
• Dataset analysis to identify representation imbalances or stereotypical correlations in training data.
• Tooling like Fairlearn, AI Fairness 360, or Aequitas to automate audits. It is the foundational diagnostic step that precedes any fairness mitigation strategy.

Algorithmic explainability refers to the suite of techniques used to make the predictions of complex, opaque models (like deep neural networks) understandable to humans. This is crucial for fairness because:

• Feature attribution methods (e.g., SHAP, LIME) reveal which input factors most influenced a specific decision, allowing auditors to check for reliance on sensitive attributes.
• Model transparency enables stakeholders to challenge and debug unfair outcomes.
• It supports right to explanation mandates in regulations like the EU's GDPR. Without explainability, diagnosing the root cause of bias is often impossible.

Differential Privacy is a rigorous mathematical framework that quantifies and bounds the privacy loss incurred by individuals when their data is used in statistical analyses or machine learning. It relates to fairness by:

• Providing a provable guarantee that the inclusion or exclusion of any single individual's data has a negligible effect on the model's output.
• Enabling the use of sensitive demographic data for bias measurement and correction without risking the exposure of private attributes.
• Allowing for the creation of privacy-preserving synthetic data for fairness testing. DP ensures privacy protections do not come at the cost of being unable to audit for bias.

Data provenance is the documented history of a dataset's origin, ownership, transformations, and processing steps. For algorithmic fairness, it provides:

• A complete audit trail to trace how potential biases were introduced, aggregated, or transformed through the data pipeline.
• Lineage tracking from raw source data to final training examples, which is critical for understanding representation issues.
• Accountability by linking model behavior back to specific data sources and processing decisions. Robust provenance is essential for reproducible fairness audits and for fulfilling regulatory compliance requirements.

Concept drift occurs when the underlying statistical relationship between the input features and the target variable that a model is trying to predict changes over time. It poses a significant challenge to long-term algorithmic fairness because:

• A model that was fair at deployment may become unfair as societal norms or real-world relationships evolve.
• Performance disparities between groups can emerge or widen silently without continuous monitoring.
• Mitigating concept drift requires continuous learning systems and fairness-aware model retraining protocols to maintain equitable performance.

Human-in-the-Loop is a system design paradigm where human judgment is integrated into an automated AI process. It is a critical mechanism for enforcing and refining algorithmic fairness:

• Edge case adjudication: Humans review and correct model predictions on ambiguous or high-stakes cases that could propagate bias.
• Bias correction: Labelers can identify and rectify stereotypical annotations in training data.
• Feedback for retraining: Human oversight generates corrected labels that are used to iteratively improve model fairness. HITL systems ensure that automated decisions remain accountable and aligned with human ethical standards.