Algorithmic Fairness

Fairness definitions are broadly categorized into group and individual notions. Group fairness (e.g., Demographic Parity, Equalized Odds) partitions a population by a sensitive attribute and requires statistical parity in outcomes or error rates across these groups. Individual fairness mandates that similar individuals receive similar predictions, formalized via a distance metric. The choice between them involves a fundamental trade-off: group metrics can mask individual injustices, while strict individual fairness can be computationally intractable and difficult to define without a task-specific similarity metric.

Bias can originate at any stage of the machine learning lifecycle, not just in the model. Key sources include:

• Historical Bias: The training data reflects existing societal inequalities.
• Representation Bias: Certain groups are under-sampled in the dataset.
• Measurement Bias: The chosen features or labels are proxies that systematically distort outcomes for a group.
• Aggregation Bias: A one-size-fits-all model fails to capture distinct subgroup dynamics.
• Evaluation Bias: The benchmark used to validate performance does not represent the target population.
• Feedback Loop Bias: A deployed model's predictions shape future data, amplifying initial biases.

Enforcing fairness constraints almost always introduces a fairness-utility trade-off, where improving a fairness metric reduces predictive accuracy or business utility. This is not a sign of failure but an inherent tension in optimization. The Pareto frontier maps the achievable combinations of accuracy and fairness. The critical task is not to eliminate the trade-off but to make an explicit, governed decision about the acceptable cost of fairness. Techniques like fairness-aware regularization allow engineers to tune this balance by adding a fairness penalty term to the loss function.

Statistical fairness metrics can be misleading because they only capture correlations. Causal fairness uses causal graphs and structural equation models to define fairness in terms of cause and effect. The gold standard is Counterfactual Fairness: a decision is fair if it would have been the same in a counterfactual world where an individual's sensitive attribute was different, while all other causally dependent features are adjusted accordingly. This approach requires strong domain knowledge to build a valid causal graph but provides a more robust, manipulation-proof definition of non-discrimination.

Algorithmic interventions to reduce bias are categorized by their point of application in the ML pipeline:

• Pre-processing: Modify the training data to remove discriminatory patterns before learning. Examples include reweighting samples and Fair Representation Learning.
• In-processing: Alter the learning algorithm itself to incorporate a fairness constraint. Examples include Adversarial Debiasing and fairness-aware regularization.
• Post-processing: Adjust the model's output predictions after training. This often involves setting group-specific decision thresholds to satisfy a metric like Equalized Odds without retraining the model.

Technical fairness must be embedded within a broader governance framework. Model Cards are standardized documents that report a model's intended use, evaluation results, and ethical considerations. Algorithmic Impact Assessments (AIAs) are structured processes to evaluate potential harms before deployment. Algorithmic Recourse ensures that individuals subjected to an unfavorable automated decision are provided with a clear, actionable path to reverse it—identifying the specific changes in their input features needed to achieve a desired outcome. These tools move fairness from a purely mathematical exercise to an accountable organizational practice.