Cauchy-Schwarz Fairness Regularizer
Keywords: Fairness, Machine Learning, Cauchy-Schwarz Divergence
TL;DR: Cauchy-Schwarz Fairness Regularizer
Abstract: In this paper, we propose a novel approach to fair machine learning, the Cauchy-Schwarz fairness regularizer, which minimizes the Cauchy-Schwarz divergence between the prediction distribution and sensitive attributes. While existing methods effectively reduce bias as indicated by low values on specific fairness metrics, they frequently struggle to achieve a balanced performance across various fairness definitions. For example, many approaches may successfully attain low demographic parity yet still demonstrate significant disparities in equal opportunity. Theoretical studies have shown that the Cauchy-Schwarz divergence provides a tighter bound compared to the Kullback-Leibler divergence and gap parity, suggesting its potential to improve fairness in machine learning models. Our empirical evaluation, conducted on four tabular datasets and one image dataset, demonstrates that the Cauchy-Schwarz fairness regularizer achieves a more balanced performance across fairness metrics while maintaining satisfactory utility. It outperforms existing fairness approaches, providing a superior trade-off between fairness and utility. In addition, the Cauchy-Schwarz fairness regularizer is a versatile, plug-and-play fairness regularizer that can be easily integrated into various machine learning models to promote fairness.
Primary Area: alignment, fairness, safety, privacy, and societal considerations
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 3986
Loading