Fairness Toolkit

The core of any toolkit is a collection of quantitative fairness metrics that mathematically define and measure bias. These metrics operationalize abstract fairness concepts into computable scores.

• Group Fairness Metrics: Calculate statistical parity between subgroups. Examples include Demographic Parity, Equal Opportunity, and Equalized Odds.
• Individual Fairness Metrics: Assess consistency between similar individuals, such as Counterfactual Fairness.
• Implementation: Libraries provide functions like demographic_parity_difference() or equal_opportunity_ratio() that take predictions, ground truth, and sensitive attributes as inputs, returning a single score (e.g., 0.15 indicates a 15% disparity).

These components automate the disaggregated evaluation of model performance to uncover hidden disparities. They move beyond aggregate metrics to slice evaluation by protected attributes.

• Disparity Visualization: Generate plots like disparity in error rates (false positive, false negative) across groups.
• Statistical Significance Testing: Use hypothesis tests (e.g., chi-squared) to determine if observed performance differences are meaningful or due to chance.
• Intersectional Analysis: Evaluate performance for subgroups at the intersection of multiple attributes (e.g., race * gender * age) to identify compounded disadvantage.
• Example: A function like MetricFrame (Fairlearn) computes accuracy, recall, F1-score for each subgroup defined by sensitive_features in a single object.

These algorithms modify the training dataset before model training to reduce underlying biases. They aim to create a fairer data distribution.

• Reweighting: Adjusts the weight of each training example to balance label distributions across groups (e.g., IBM AIF360's Reweighting).

• Disparate Impact Remover: A massaging technique that edits feature values to reduce correlation with protected attributes while preserving rank-ordering within groups.

• Learning Fair Representations: An optimization method that transforms data into a new, latent representation where it's difficult to predict the protected attribute from the encoded features, while preserving utility for the main task.

• Use Case: Applied when you have a biased historical dataset and want to train a standard model (e.g., logistic regression) on a corrected version.

These techniques modify the model training process itself by incorporating fairness as a constraint or objective directly into the learning algorithm.

• Adversarial Debiasing: Uses a minimax game where a predictor model tries to be accurate, while an adversary model tries to predict the protected attribute from the predictor's embeddings. This decorrelates the internal representations from sensitive attributes (e.g., IBM AIF360's AdversarialDebiasing).

• Fairness-Constrained Optimization: Adds a mathematical fairness constraint (e.g., demographic parity) as a penalty term to the model's loss function. The ExponentiatedGradient reduction in Fairlearn reduces fair classification to a sequence of cost-sensitive classification problems.

• Use Case: Essential when you need to train a new model from scratch with fairness baked into its core parameters.

These methods adjust a trained model's predictions or decision thresholds after inference to satisfy a fairness criterion, without retraining the model.

• Threshold Optimizer: Finds group-specific decision thresholds that achieve a target fairness metric (e.g., equalized odds) with minimal impact on overall accuracy. This is implemented in Fairlearn's ThresholdOptimizer.

• Reject Option Classification: For instances where the model's prediction confidence is low (near the decision boundary), the outcome is assigned to favor the disadvantaged group.

• Advantage: Highly practical for deployed models, as it requires only the output scores and sensitive attributes, not model internals or training data.

• Limitation: Requires knowing the sensitive attribute at inference time, which may not always be permissible.

Tools to document, visualize, and communicate fairness assessments to technical and non-technical stakeholders, ensuring transparency and auditability.

• Bias Audit Reports: Automatically generate summary reports detailing metrics, visualizations, and mitigation results across different fairness definitions.

• Model Cards for Fairness: Extend standard model cards with dedicated sections for fairness performance, including disaggregated evaluation tables, known trade-offs, and recommended mitigation strategies.

• Interactive Dashboards: Allow users to explore trade-offs between different fairness metrics and overall accuracy (e.g., Fairlearn's FairnessDashboard).

• Example: The FairlearnDashboard widget lets users visualize disparities and interactively apply post-processing thresholds to see the resulting fairness/accuracy trade-off curve.