Robust Algorithmic Recourse Design Under Model Shifts

Yulu Jin, Lifeng Lai

Keywords: Algorithmic recourse, Model shifts, Robustness, Uncertainty quantification

Abstract: Algorithmic recourse offers users recommendations for actions that can help alter unfavorable outcomes in practical decision-making systems. Although many methods have been proposed to design easily implementable recourses, model updates or shifts may render previously generated recourses invalid. To assess the robustness of recourses against model shifts, we propose an uncertainty quantification method to calculate a theoretical upper-bound of the recourse invalidation rate for any counterfactual plan and any prediction model, without requiring distributional assumptions about the feature space. Furthermore, given the inherent trade-off between recourse cost and recourse robustness, users should be empowered to manage the implementation cost versus robustness trade-off. To this end, we propose a novel framework that leverages the derived invalidation rate bounds to generate model-agnostic recourses that satisfy the user's specified invalidation needs. Numerical results on multiple datasets demonstrate the effectiveness of the derived theoretical bounds and the efficacy of the proposed algorithms.

Primary Area: probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)

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: 4220