XAI Procedural Fairness Auditing Framework: avoid misguided outcomes by refocusing on fairness properties
Keywords: Interpretable Machine Learning, Explainable Artificial Intelligence, Trustworthy AI, Fair ML
TL;DR: We present a framework for XAI auditing through a procedural fairness lens as applied to the XAI surrogate generation algorithm, STEALTH.
Abstract: Interpretable machine learning and explainable AI (XAI) methods used to investigate fairness properties can be described as ML auditing. Current ML researchers have noted that there are limited, successful implementations of procedural fairness, which focuses on the decision-making steps rather than fair outcomes. We present the results of our procedural fairness auditing framework for XAI tools. We evaluated STEALTH, an ensemble XAI method that combines novel global surrogate model generation that avoids detection by deceptive models with well-known LIME's local explanations. Through a Procedural Fairness lens, we audited STEALTH's decision-making process outside of its notable performance outcomes. The procedural fairness audit reports that STEALTH's global surrogate models are impressive and a successful application of recursive bi-clustering for representative data downsampling. However, the audit also revealed STEALTH's training data biases, and we discuss how STEALTH's fairness claims were misguided by ``fairer outcomes.'' The procedural fairness auditing framework provides an outline of how to interpret ML decision-making, ensuring procedural fairness.
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.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/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: 6300
Loading