Amicable Perturbations

Jesse Friedbaum, Ravi Tandon

Primary Area: general machine learning (i.e., none of the above)

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: Machine Learning, Robust Learning, Counterfactual Explanations, AI Decision Making

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.

TL;DR: We propose a new framework to find the most efficient way to change an undesirable classification.

Abstract: Machine learning based classifiers have achieved incredible success in a variety of sectors such as college admissions, hiring and banking. However their ability to make classifications has not been fully exploited to understand how to improve undesirable classifications. We propose a new framework for finding the most efficient changes that could be made in the real world to achieve a more favorable classification, and term these changes \textit{amicable perturbations}. We present a principled methodology for creating amicable perturbations and demonstrate their effectiveness on data sets from a variety of fields. Amicable perturbations differ from counterfactuals in that they are better suited to balance the effort-reward trade-off and lead to the most efficient plan of action. Unlike adversarial examples, which fool a classifier into making false prediction, amicable perturbations are intended to affect the true class of the data point. To this end, we develop a novel method for verifying that a amicable perturbations change the true class probabilities. We also compare our results to those achieved by previous methods such as counterfactuals and adversarial attacks.

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: 339