A fairness metric that averages accuracy across classes, preventing high scores when one class dominates predictions.