FairnessConstraints

FairnessConstraints refer to a set of mathematical restrictions imposed on a learning problem to ensure that model predictions satisfy predefined fairness criteria across protected attribute groups. They are used to formalize obligations such as equal treatment or equal impact for different demographic groups during model training and evaluation.

Common fairness notions expressed by such constraints include demographic parity, equalized odds, and equal opportunity. Demographic parity requires equal prediction rates across groups, regardless of actual outcomes. Equalized odds demands equal false positive and true positive rates across groups. Equal opportunity focuses on equal true positive rates. In addition to group-based notions, individual fairness aims for similar individuals to receive similar predictions, though this is harder to express with simple group constraints. FairnessConstraints are typically written as equalities or inequalities on predicted outcomes, scores, or error rates, for example restricting the difference in average predictions between groups to be within a tolerance.

In practice, fairness constraints are incorporated through in-processing methods, which modify the learning objective to include the constraints, often via Lagrangian multipliers or penalty terms. They can also be used in pre-processing steps or post-processing adjustments, but in-processing remains a common approach for optimizing both accuracy and fairness simultaneously.

Challenges include trade-offs between predictive accuracy and fairness, choice of fairness notion appropriate to the context, and availability of sensitive attribute information. Additionally, satisfying one notion may violate another, and some settings may render the constraints infeasible. FairnessConstraints thus require careful specification, context-aware selection of fairness goals, and ongoing evaluation.