On Demographic Group Fairness Guarantees in Deep Learning
- Published
- Record updated
Summary
The authors present a theoretical framework that derives bounds on fairness errors and convergence rates, accounting for how data distributions differ across demographic groups. They validate it on six datasets spanning image, tabular and text data, finding that feature distribution differences significantly affect model fairness, especially across racial categories. They also propose Fairness-Aware Regularization (FAR), a training objective that minimizes inter-group discrepancies in feature centroids and covariances, and report consistent improvements in overall AUC, ES-AUC and subgroup performance.
Mitigation
Fairness-Aware Regularization (FAR), a training objective that directly minimizes inter-group discrepancies in feature centroids and covariances.