Measuring Fairness Using Probable Segmentation for Continuous Sensitive Attributes
Primary Area: societal considerations including fairness, safety, privacy
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.
Keywords: Fairness, Probable Demographic Parity, Continuous Sensitive Attribute
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.
TL;DR: This study introduces a fairness metric called probable demographic parity for continuous sensitive attributes.
Abstract: Algorithmic fairness in machine learning aims to regulate the bias towards sensitive attributes. In the case of continuous sensitive attributes, however, defining and measuring fairness is a non-trivial task. For instance, estimating a maximum disparity of predictions within a continuous sensitive attribute is vulnerable for an extreme case, whereas a mean disparity of predictions underestimates the effect of the worst case, which is meaningful for testing the independence between the prediction and the sensitive attribute. We address this issue by developing a new definition of fairness, probable demographic parity, based on a maximum prediction disparity of probable segments. We only consider probable segments of the continuous sensitive attribute that have a higher probability than the preset minimum probability condition. Then, we compare the local prediction average of these segments to identify the maximum prediction disparity. By doing so, we ensure a consistent estimation error for the local prediction average of the segment and mitigate the risk of encountering missing data in the segment. We analyze the various theoretical features including stability and independence and experimentally prove the usefulness of the proposed metric.
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: 3691
Loading