Language Models Can Help to Learn High-Performing Cost Functions for Recourse

Eoin M. Kenny, Allan Anzagira, Tom Bewley, Freddy Lecue, Manuela Veloso

Keywords: algorithmic recourse, large language models, cost functions, interpretable ml, user study

TL;DR: We show that LLMs can help to train good cost functions for recourse.

Abstract:

Algorithmic recourse is a specialised variant of counterfactual explanation, concerned with offering actionable recommendations to individuals who have received adverse outcomes from automated systems. Most recourse algorithms assume access to a cost function, which quantifies the effort involved in following recommendations. Such functions are useful for filtering down recourse options to those which are most actionable. In this study, we explore the use of large language models (LLMs) to help label data for training recourse cost functions, while preserving important factors such as transparency, fairness, and performance. We find that LLMs do generally align with human judgements of cost and can label data for the training of effective cost functions, moreover they can be fine-tuned with simple prompt engineering to maximise performance and improve current recourse algorithms in practice. Previously, recourse cost definitions have mainly relied on heuristics and missed the complexities of feature dependencies and fairness attributes, which has drastically limited their usefulness. Our results show that it is possible to train a high-performing, interpretable cost function by consulting an LLM via careful prompt engineering. Furthermore, these cost functions can be customised to add or remove biases as befitting the domain and problem.
Overall, this study suggests a simple, accessible method for accurately quantifying notions of cost, effort, or distance between data points that correlate with human intuition, with possible applications throughout the explainable AI field.

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: 7915