Historical Bias

A data-level flaw where the training sample fails to accurately reflect the target population. While historical bias reflects real-world structural inequalities, representation bias is a sampling error that exacerbates these issues.

• Under-sampling: Minority groups appear too infrequently to learn robust patterns.
• Over-sampling: Dominant groups drown out minority signals.
• Labeling disparity: Annotation quality varies systematically across groups.

A model trained on a dataset with both historical and representation bias will compound errors, producing predictions that are doubly distorted.

The systematic study of designing machine learning systems that make decisions without unjustified discrimination. While historical bias describes a problem in the data, algorithmic fairness provides the mathematical frameworks to measure and constrain it.

• Group fairness: Statistical parity, equalized odds, and equal opportunity.
• Individual fairness: Similar individuals should receive similar outcomes.
• Causal fairness: Uses structural causal models to isolate discriminatory pathways.

Fairness metrics quantify the downstream impact of historical bias, but no single metric captures all ethical dimensions.

A legal doctrine and statistical measure identifying facially neutral policies that disproportionately harm protected groups. The Four-Fifths Rule serves as a practical threshold: a selection rate for any group below 80% of the highest group's rate indicates potential adverse impact.

• Originates from U.S. employment discrimination law.
• Applies to algorithmic decisions even when protected attributes are excluded.
• Historical bias in training data is a primary driver of disparate impact in model outputs.

Proving disparate impact requires rigorous statistical testing across demographic segments.

Technical interventions applied at three stages of the machine learning pipeline to reduce unwanted algorithmic bias stemming from historical bias in data.

• Pre-processing: Reweighting, resampling, or transforming training data to remove discriminatory patterns before training.
• In-processing: Adding fairness constraints directly into the model's objective function during training.
• Post-processing: Adjusting model outputs or decision thresholds after prediction to satisfy fairness criteria.

No single technique universally resolves historical bias; the choice depends on the fairness definition, regulatory context, and operational constraints.

A causal definition of fairness where a decision is fair if it remains the same in a counterfactual world where the individual belonged to a different demographic group. This approach directly addresses historical bias by modeling the causal pathways through which protected attributes influence outcomes.

• Requires a structural causal model of the data-generating process.
• Distinguishes between discriminatory path-specific effects and legitimate explanatory variables.
• Example: If changing only an applicant's race (while holding all causally downstream variables constant) changes the loan decision, the model is counterfactually unfair.

This framework is powerful but computationally demanding and requires strong causal assumptions.

A framework evaluating algorithmic bias across overlapping social identities. Historical bias often compounds at intersections—a Black woman may face discrimination that is not captured by examining race or gender in isolation.

• Kimberlé Crenshaw's foundational work demonstrated that single-axis analysis erases those at the intersection.
• Standard fairness metrics applied to single protected attributes can mask severe subgroup disparities.
• The Gender Shades study revealed that commercial facial recognition error rates for darker-skinned women were up to 34% higher than for lighter-skinned men.

Intersectional auditing requires disaggregated evaluation across all combinations of protected attributes.