Fast Conditional Intervention in Algorithmic Recourse with Reinforcement Learning

Chiung-Yu Chang, Keng-Te Liao, Mi-Yen Yeh, Shou-De Lin

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: Algorithmic recource, Causality, Reinforcement Learning, Explainable machine learning

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

Abstract: Explaining the decisions made by machine learning classifiers aids individuals in identifying critical factors and charting future plans. Recent studies have shown that incorporating causal graphs of input features provides more realistic explanations; however, this also introduces new challenges such as handling noisy graphs and efficiently performing inference with black-box classifiers. In this work, we tackle these issues by presenting an efficient reinforcement learning (RL)-based approach with an idea of conditional intervention. Our intervention method is theoretically preferable and considers both feature dependencies and incompleteness of graphs. Simultaneously, the RL-based method offers the capacity to learn the intervention process while guarantees computational complexity at inference stage. In the experiments, we showcase the efficiency and superior performance of our solution when compared to baseline methods on both synthetic and real datasets.

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: 9061