ReLax: A Recourse Explanation Library in Jax
Open Access
- Author:
- Xiong, Xinchang
- Area of Honors:
- Computer Science
- Degree:
- Bachelor of Science
- Document Type:
- Thesis
- Thesis Supervisors:
- Amulya Yadav, Thesis Supervisor
Danfeng Zhang, Thesis Honors Advisor
C Lee Giles, Thesis Supervisor - Keywords:
- Algorithmic Recourse
ExplainableArtificial Intelligence
Interpretability
Counterfactual Explanation
Algorithmic Recourse
Explainable Artificial Intelligence - Abstract:
- Counterfactual explanation has been proposed by several works as a means to explain the decisions of machine learning models. By providing actionable suggestions to achieve the desired outcome, counterfactual explanation enables affected end users to better understand the decision-making process. As the number of counterfactual explanation methods increases, there is a pressing need for a standardized benchmarking platform that allows researchers to compare and assess prior works using standardized evaluation metrics. However, there are currently limited options for benchmarking, with existing solutions relying on solving optimization problems for each input data point iteratively, making them not scalable. Addressing these limitations, we present ReLax (Recourse Explanation Library in Jax), a jax-based Python library designed for benchmarking counterfactual explanation methods on different datasets and developing new counterfactual explanation methods. In summary, our work provides the following contributions: (i) fast and scalable benchmarking on 4 state-of-art counterfactual explanation methods, (ii) ReLax provides a comprehensible and easy-to-use API for users to customize and develop new counterfactual explanation methods, (iii) a standardized evaluation metrics and datasets for transparent and extensive comparisons of these methods. Our extensive experiments on multiple real-world datasets show that ReLax runs 134.37 times faster on average compared to CARLA a benchmark library for CF explanations.