Bias Mitigation

Techniques applied to the training data before model fitting to remove discriminatory patterns.

• Reweighting: Assigns higher weights to underrepresented group instances to balance the dataset.
• Disparate Impact Removal: Transforms feature values to break the statistical dependency between protected attributes and outcomes.
• Data Augmentation: Synthetically generates samples for minority groups to correct representation bias.

Example: The Optimized Pre-Processing algorithm in AIF360 edits feature values and labels simultaneously to satisfy fairness constraints while minimizing data distortion.

Fairness constraints integrated directly into the model training objective through regularization or adversarial learning.

• Adversarial Debiasing: Trains a predictor network to maximize accuracy while an adversary network attempts to predict the protected attribute from the predictor's output. The predictor is penalized when the adversary succeeds.
• Prejudice Remover: Adds a fairness-aware regularization term to the loss function that penalizes mutual information between predictions and protected attributes.
• Exponentiated Gradient Reduction: Treats fairness as a constrained optimization problem, finding the classifier that minimizes error subject to equalized odds or demographic parity constraints.

Adjustments applied to model outputs after prediction to achieve fairness without retraining the underlying classifier.

• Threshold Optimization: Sets different decision thresholds for each protected group to equalize true positive rates or false positive rates.
• Calibrated Equalized Odds: Derives group-specific probability thresholds from a held-out calibration set to satisfy equalized odds while preserving calibration.
• Reject Option Classification: Introduces a band of uncertainty around the decision boundary where predictions for privileged groups are flipped to favorable outcomes and predictions for unprivileged groups are flipped to unfavorable outcomes.

Key advantage: These methods are model-agnostic and can be applied to any black-box classifier.

The fundamental tension between enforcing strict fairness constraints and maintaining predictive accuracy.

• Imposing demographic parity often reduces overall accuracy because it forces the model to ignore legitimate predictive features correlated with protected attributes.
• Equalized odds typically preserves more utility than demographic parity since it conditions on the true outcome.
• The Pareto frontier of fairness-utility trade-offs can be explored using multi-objective optimization to let stakeholders choose an acceptable operating point.

Practical guidance: Always measure the cost of fairness in terms of business metrics, not just abstract error rates, before selecting a mitigation strategy.

Methods that use structural causal models to distinguish discriminatory path-specific effects from legitimate influences.

• Counterfactual Fairness: A decision is fair if it would remain the same in a counterfactual world where the individual's protected attribute were different, holding all other causally independent factors constant.
• Path-Specific Effect Removal: Identifies and neutralizes only the causal pathways from protected attributes to outcomes that are deemed discriminatory, while preserving pathways through resolving variables like education or experience.
• Fair Inference on Outcomes: Uses causal graphs to determine which observed variables can be legitimately used for prediction without propagating historical discrimination.

These approaches require explicit domain knowledge to construct the causal graph.

A systematic framework for verifying that bias mitigation has succeeded without introducing new harms.

• Disaggregated Evaluation: Measure fairness metrics separately for each intersectional subgroup, not just aggregate protected attributes.
• Minimum Acceptable Performance: Define a floor for model accuracy below which mitigation is considered too costly to deploy.
• Robustness Testing: Evaluate whether fairness improvements generalize to distribution shifts and out-of-sample data.
• A/B Deployment: Run the original and mitigated models in parallel to compare real-world outcomes before full rollout.

Tools: Fairlearn and AIF360 provide built-in assessment dashboards for this validation workflow.

Trains a primary classifier to predict the target variable while a simultaneous adversarial network attempts to predict the protected attribute from the primary model's output. The primary model is penalized for encoding sensitive information.

• Architecture: Dual-network gradient reversal layer
• Protects: Race, gender, age, and other legally defined attributes
• Outcome: Learns a latent representation that is maximally informative for the task but minimally informative for discrimination
• Key insight: Fairness is learned as a side-effect of adversarial pressure

Uses a Structural Causal Model (SCM) to simulate what a decision would have been if an individual's protected attribute had been different, holding all non-discriminatory factors constant.

• Framework: Pearl's do-calculus for causal inference
• Process: Generates a counterfactual world, re-evaluates the decision, and measures divergence
• Strength: Distinguishes legitimate causal paths from discriminatory ones
• Challenge: Requires a correctly specified causal graph, which is domain-expertise intensive

Quantifies adverse impact using the Four-Fifths Rule: the selection rate for any protected group must be at least 80% of the rate for the group with the highest selection rate.

• Formula: Impact Ratio = Selection Rate (Protected) / Selection Rate (Reference)
• Threshold: Ratio < 0.80 triggers a bias audit
• Origin: U.S. Uniform Guidelines on Employee Selection Procedures (1978)
• Modern use: Embedded in model cards and fairness dashboards for continuous monitoring