Fairness of Classifiers with Feature Constraints

A new arXiv paper (2605.00592) proposes a definition of fairness for machine learning classifiers that accounts for constraints between features. The authors argue that a decision is fair if it has a fair explanation—a prime-implicant reason containing no protected features, with constraints considered. They show that ignoring constraints can alter fairness judgments even without direct links between protected and unprotected features. Three classifier-level fairness definitions are introduced: all decisions have only fair explanations, at least one fair explanation exists, or changing explanations affects fairness.