Fairly Explaining Monotonic Models: a New Shapley Value

Dangxing Chen, Jingfeng Chen

Keywords: Attribution problem, Model interpretation, Shapley value, Fairness

TL;DR: We propose a new Shapley value for fair explanation of monotonic models that differs from the classical Shapley value

Abstract: The Shapley value has been widely used as an attribution method for explaining black-box machine learning models. A rigorous mathematical framework based on a number of axioms has enabled Shapley value to disentangle the black-box structure of models. Recent studies have shown that domain knowledge is an important component of machine learning models. Science-informed machine learning models that incorporate domain knowledge have demonstrated better generalization and interpretation capabilities. But do we obtain consistent scientific explanations when we apply attribution methods to science-informed machine learning models? In this study, we show that Shapley value cannot be guaranteed to reflect domain knowledge, such as monotonicity. To remedy Shapley's monotonicity failure, we propose a new version of Shapley value. As a result of extensive analytical and empirical examples, we show that Shapley value often produces misleading explanations for monotonic models, which can be avoided using the new method.

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.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/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: 1613