Feasible Algorithmic Recourse Without Explicit Structure Prior
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: Deep learning, Black box, Algorithmic Recourse, Interpretability, Feasibility
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.
TL;DR: This paper proposes an approach to learning feasible algorithmic recourse without strong prior knowledge.
Abstract: To ensure that vulnerable end-users have a clear understanding of decisions made by black-box models, algorithmic recourse has made significant progress by identifying small changes in input features that can alter predictions. However, the recoursed examples in real-world scenarios are only feasible and actionable for end-users if they preserve the realistic constraints among input features. Previous works have highlighted the importance of incorporating causality into algorithmic recourse to capture these constraints as causal relationships. Existing methods often rely on inaccessible prior Structural Causal Models (SCMs) or complete causal graphs. To maintain the causal relationships without such prior knowledge, we propose a novel approach that focuses on identifying and constraining the variation of exogenous noise by leveraging recent advancements in non-linear Independent Component Analysis (ICA). Based on this idea, we introduce two methods: Algorithmic Recourse with L2 norm (AR-L2) and Algorithmic Recourse with Nuclear norm (AR-Nuc). Experimental results on synthetic, semi-synthetic, and real-world data demonstrate the effectiveness of our proposed methods.
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: 7165
Loading