AI Fairness 360

AI Fairness 360 implements a taxonomy of bias mitigation interventions across the AI lifecycle:

• Pre-processing: Reweighing and Optimized Preprocessing transform training data to remove discrimination before modeling
• In-processing: Adversarial Debiasing and Prejudice Remover add fairness constraints directly into the model's objective function during training
• Post-processing: Equalized Odds Postprocessing and Calibrated Equalized Odds adjust model predictions after training to satisfy fairness criteria

The toolkit provides a unified API so practitioners can swap mitigation strategies without rewriting their pipeline.

The toolkit exposes a comprehensive library of fairness metrics to quantify bias from multiple mathematical perspectives:

• Group Fairness: Statistical Parity Difference, Disparate Impact Ratio, Average Odds Difference, and Equal Opportunity Difference measure outcome disparities between privileged and unprivileged groups
• Individual Fairness: Consistency metrics evaluate whether similar individuals receive similar predictions regardless of group membership
• Generalized Entropy Index: Captures inequality across the entire distribution, decomposable into between-group and within-group components

Each metric returns a scalar value and a Boolean flag indicating whether a user-defined fairness threshold has been violated.

AI Fairness 360 introduces a custom StructuredDataset abstraction that explicitly encodes metadata about which features are:

• Protected attributes: Columns containing sensitive characteristics such as race, gender, or age that define privileged and unprivileged groups
• Favorable labels: The positive outcome class that fairness metrics evaluate for equitable distribution
• Instance weights: Optional sample-level weights for handling imbalanced representations

This structured wrapper ensures that every algorithm and metric in the toolkit correctly interprets the ethical semantics of the data, preventing silent misapplication of fairness definitions.

Beyond detection and mitigation, the toolkit connects to explainability modules that help practitioners understand why a model exhibits bias:

• Metric Explainer: Decomposes a fairness metric into feature-level contributions, identifying which input variables drive disparate outcomes
• Disparate Impact Remover explainer: Visualizes how the pre-processing transformation altered the joint distribution of features and labels
• JSON serialization: All explanations are exportable for audit trails and governance reporting

This bridges the gap between identifying a fairness violation and diagnosing its root cause in the data or model architecture.

The toolkit ships with curated benchmark datasets that have become standard references in the fairness research community:

• Adult Census Income: Predicts whether income exceeds $50K/year, with sex and race as protected attributes
• COMPAS Recidivism: Criminal risk assessment scores with documented racial disparities
• German Credit: Creditworthiness classification with age and gender fairness considerations
• Bank Marketing: Term deposit subscription prediction with marital status as a sensitive feature

Each dataset is pre-processed with defined privileged and unprivileged groups, enabling reproducible fairness research and consistent benchmarking across mitigation techniques.