Bias Mitigation

Methods applied to the training dataset before model training to remove underlying biases. The goal is to create a fairer data distribution.

• Reweighting: Adjusts the importance (weight) of individual data samples to balance outcomes across protected groups.
• Disparate Impact Remover: A statistical technique that edits feature values to reduce correlation with protected attributes while preserving rank ordering.
• Optimized Pre-processing: Learns a probabilistic transformation of the data to maximize fairness and utility simultaneously.

These techniques treat the data as the source of bias, aiming to correct historical inequities before they are learned by the model.

Methods applied during the model training process by modifying the learning algorithm itself to incorporate fairness objectives.

• Adversarial Debiasing: A neural network architecture where a primary predictor is trained for accuracy while an adversarial network is trained to prevent prediction of the protected attribute from the primary model's representations.
• Fairness Constraints: Mathematical conditions (e.g., demographic parity, equalized odds) are added directly to the model's loss function as regularization terms.
• Reductions Approach: Converts fairness constraints into a sequence of weighted classification problems, allowing standard algorithms to be used.

In-processing provides direct control over the trade-off between accuracy and fairness during optimization.

Methods applied to a trained model's output predictions to achieve fairness, without modifying the model or training data. This is often the most practical approach for deployed systems.

• Threshold Optimization: Adjusts the decision threshold (the cutoff for a positive prediction) independently for different demographic groups to meet a target fairness metric like equal opportunity.
• Rejection Option Classification: Withholds predictions (returns "reject") for instances where the model's confidence is low and the predicted label differs from a fairness-aware alternative label.

Post-processing is model-agnostic and useful for achieving fairness after a model is already in production, though it requires access to protected attributes at inference time.

A specific and powerful in-processing technique that uses a game-theoretic setup. A primary model (the predictor) is trained to perform the main task (e.g., loan approval). Simultaneously, an adversarial model is trained to predict the protected attribute (e.g., gender) from the primary model's internal representations (e.g., hidden layer activations).

The primary model's objective becomes twofold: maximize task accuracy while minimizing the adversarial model's ability to predict the protected attribute. This forces the primary model to learn representations that are invariant with respect to the protected attribute, thereby removing bias from its decision process. It is particularly effective for deep neural networks.

Mathematical formulations of fairness goals that are integrated into a model's training objective. Instead of just minimizing prediction error, the model also minimizes a fairness penalty.

Common constraint types include:

• Demographic Parity: P(Ŷ=1 | A=a) = P(Ŷ=1 | A=b) for all groups a, b. The selection rate is equal.
• Equalized Odds: P(Ŷ=1 | A=a, Y=y) = P(Ŷ=1 | A=b, Y=y) for all a, b and *y ∈ {0,1}`. Equal true positive and false positive rates.
• Equal Opportunity: A relaxed form of equalized odds requiring only true positive rates to be equal.

These constraints are often enforced via Lagrangian multipliers, turning the constrained optimization problem into an unconstrained one the model can solve.

Advanced techniques that use causal inference frameworks to define and achieve fairness. They move beyond statistical correlations to model cause-and-effect relationships.

• Counterfactual Fairness: A model is considered fair if its prediction for an individual is the same in the actual world and in a counterfactual world where that individual's protected attribute (e.g., race) had been different, holding all else equal. This requires a causal model of the data-generating process.
• Path-Specific Counterfactuals: Extends counterfactual fairness to analyze and mitigate bias transmitted through specific causal pathways (e.g., via education level but not via neighborhood).

These methods provide a strong, individual-level notion of fairness but require significant domain knowledge to specify the correct causal graph.

Algorithmic fairness is the study and application of principles to ensure automated systems do not create unjust outcomes based on protected attributes like race or gender. It moves from abstract ethics to concrete, measurable engineering goals.

• Core Challenge: Translating qualitative notions of 'fairness' into quantitative fairness metrics that can be optimized or constrained.
• Key Tension: Often exists between fairness and pure predictive accuracy, requiring trade-off analysis.
• Implementation: Achieved through formal fairness constraints during model training or via post-hoc adjustments to predictions.

A bias audit is a systematic, documented evaluation of an AI system to detect and measure discriminatory bias. It is a foundational governance activity, often required by regulations like the NYC AI Bias Law.

• Process: Involves subgroup analysis to calculate performance metrics (e.g., accuracy, FPR) for each protected group.
• Outputs: Produces a report quantifying disparate impact (e.g., a 20% lower approval rate for a protected group).
• Tools: Conducted using fairness toolkits like AIF360 or Fairlearn, which provide standardized metrics and tests.

Pre-processing mitigation techniques modify the training data before model training to remove underlying biases. The goal is to create a 'fairer' dataset.

• Reweighting: Adjusts the importance of samples from different groups to balance influence during training.
• Disparate Impact Remover: A technique that edits feature values to reduce correlation with protected attributes while preserving rank ordering.
• Use Case: Ideal when you have control over the data pipeline and want to address historical bias at its source. It is model-agnostic.

In-processing mitigation modifies the model training algorithm itself to incorporate fairness directly into the learning objective. It bakes fairness into the model's parameters.

• Adversarial Debiasing: The primary model learns to make accurate predictions while an adversarial network tries to predict the protected attribute from its internal representations, forcing them to be non-discriminatory.
• Constrained Optimization: The model is trained to maximize accuracy subject to a fairness constraint like demographic parity or equalized odds.
• Use Case: Provides a direct, integrated approach but is often specific to a model family and fairness definition.

Post-processing mitigation adjusts a trained model's predictions after they are made to satisfy a fairness criterion. It treats the model as a fixed 'black box'.

• Threshold Adjustment: Applies different classification thresholds to different demographic groups to equalize metrics like equal opportunity.
• Advantage: Does not require retraining the model or access to the training pipeline. Fast to implement and test.
• Limitation: Requires knowing the protected attribute at inference time, which can raise privacy concerns. It may also reduce overall accuracy.

These are two legal and technical categories of algorithmic bias.

• Disparate Treatment: Explicit discrimination. Occurs when a model directly uses a protected attribute (e.g., 'gender' as an input feature) to make different decisions. This is often easier to detect and remediate by removing the feature.

• Disparate Impact: Implicit discrimination. Occurs when a model uses a proxy variable (e.g., 'zip code' highly correlated with race) or patterns in other features that result in disproportionately adverse outcomes for a protected group. This is more insidious and requires techniques like bias auditing to uncover.