Everyone Deserves Recourse: Feasible Recourse Paths Using Data Augmentation

Shubham Sharma, Leonidas Tsepenekas, Margarita Boyarskaya, Freddy Lecue

Keywords: Recourse, Explainability

TL;DR: A distance and threshold learning framework combined with augmentation to provide multi-step recourse paths to individuals getting negative decisions from AI models

Abstract:

Decisions made using machine learning models can negatively impact individuals
in critical applications such as healthcare and finance by denying essential services
or access to opportunity. Algorithmic recourse supplements a negative AI decision
by providing rejected individuals with advice on the changes they can make to their
profiles, so that they may eventually achieve the desired outcome. Most existing
recourse methods provide single-step changes by using counterfactual explanations.
These counterfactual explanations are computed assuming a fixed (not learned)
distance function. Further, few works consider providing more realistic multi-step
changes in the form of recourse paths. However, such methods may fail to provide
any recourse path for some individuals or provide paths that might not be feasible,
since intermediate steps needed to reach the counterfactual explanation may not
be realizable. We introduce a framework for learning an optimal distance function
and threshold to compute multi-step recourse paths for all. First, we formalize the
problem of finding multi-step recourse paths. Given a set of feasible transitions, we
propose a data-driven framework for learning the optimal distance and threshold
for each step with PAC (Probably Approximately Correct) guarantees. Finally,
we provide a data augmentation algorithm to ensure that a solution exists for all
individuals. Experiments on several datasets show that the proposed method learns
feasible recourse paths for all individuals.

Primary Area: interpretability and explainable AI

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.

Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.

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: 12043