Year: 2024
Type: article
Abstract: From healthcare to criminal justice, machine learning (ML) models have permeated society to support domain experts in making decisions.Given the high-stakes nature of decision outcomes in some real-world domains, concerns over the trustworthiness of ML model predictions have raised increasing attention.These concerns have spurred surging research interests in explainable artificial intelligence (XAI), whose mission is to equip end-users with an understanding (or explanation) of AI decision making, and to help provide assessments to end-users about when to rely on ML models and when to exercise caution.Within the XAI domain, recourse 1 has emerged as a notable technique, which provides alternative scenarios (which lead to desirable AI decisions) to individuals adversely affected by ML predictions, thereby elucidating the underlying decision-making mechanisms to end users.For instance, recourse methods can provide corrective suggestions for loan applicants who have been rejected by a bank's ML algorithm, or give practical advice to teachers handling students at risk of dropping out from school.Numerous recourse explanation methods have been recently proposed.Yet, the substantial runtime overhead imposed by many recourse explanation methods compels current research to limit their evaluation and benchmarking to medium-sized datasets (i.e., ~50k data points).This limitation has significantly impeded progress in the field of algorithmic recourse, while it also raises valid concerns about the scalability of existing approaches.To address this challenge, we propose ReLax, a JAX-based benchmarking library, designed for efficient and scalable recourse generation.ReLax supports a variety of recourse methods and datasets, demonstrating performance improvements of at least two orders of magnitude over current libraries.Notably, ReLax can benchmark real-world datasets up to 10 million data points, a 200-fold increase over existing norms, without imposing prohibitive computational costs. Statement of needRecourse and counterfactual explanation methods concentrate on the generation of new instances that lead to contrastive predicted outcomes (Karimi et al., 2022;Stepin et al., 2021;Verma et al., 2020).Given their ability to provide actionable recourse, these explanations are often favored by human end-users (Bhatt et al., 2020;Binns et al., 2018;Miller, 2019).Despite progress made in counterfactual explanation research (
DOI: 10.21105/joss.06567
Authors: Hangzhi Guo, Xinchang Xiong, Wenbo Zhang, Amulya Yadav
Institutions: Pennsylvania State University, Duke University
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: diamond
OpenAlex ID: W4404296731