Learning Actionable Counterfactual Explanations in Large State Spaces

Keziah Naggita, Matthew Walter, Avrim Blum

Keywords: counterfactual explanations, recourse, data-driven algorithms, fairness

TL;DR: We propose three data-driven CFE generators that use data beyond that key to classification to produce generalizable CFEs desirable to decision-makers and individuals.

Abstract:

An increasing number of high-stakes domains rely on machine learning to make decisions that have significant consequences for individuals, such as in loan approvals and college admissions. The black-box nature of these processes has led to a growing demand for solutions that make individuals aware of potential ways they could improve their qualifications.
Counterfactual explanations (CFEs) are one form of feedback commonly used to provide insight into decision-making systems. Specifically, contemporary CFE generators provide explanations in the form of low-level CFEs whose constituent actions precisely describe how much a negatively classified individual should add to or subtract from their input features to achieve the desired positive classification.
However, the low-level CFE generators have several shortcomings: they are hard to scale, often misaligned with real-world conditions, constrained by information access (e.g., they can not query the classifier), and make inadequate use of available historical data.
To address these challenges, we propose three data-driven CFE generators that create generalizable CFEs with desirable characteristics for individuals and decision-makers.
Through extensive empirical experiments, we compare the proposed CFE generators with a low-level CFE generator on four real-world (BRFSS, Foods, and two NHANES datasets), five semi-synthetic, and five variants of fully-synthetic datasets.
Our problem can also be seen as learning an optimal policy in a family of large but deterministic Markov decision processes.

Primary Area: interpretability and explainable AI

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/2025/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: 7978