A quantitative approach for the comparison of additive local explanation methods

Local additive explanation methods are increasingly used to understand the predictions of complex Machine Learning (ML) models. The most used additive methods, SHAP and LIME, suffer from limitations that are rarely measured in the literature. This paper aims to measure these limitations on a wide range (304) of OpenML datasets using six quantitative metrics, and also evaluate emergent coalitional-based methods to tackle the weaknesses of other methods. We illustrate and validate results on a specific medical dataset, SA-Heart. Our findings reveal that LIME and SHAP’s approximations are particularly efficient in high dimension and generate intelligible global explanations, but they suffer from a lack of precision regarding local explanations and possibly unwanted behavior when changing the method’s parameters. Coalitional-based methods are computationally expensive in high dimension, but offer higher quality local explanations. Finally, we present a roadmap summarizing our work by pointing out the most appropriate method depending on dataset dimensionality and user’s objectives. •A methodology to compare local explanation methods is proposed (including new metrics).•Machine Learning models complexity have an impact on explanations.•Additive local explanation methods are complementary.•Trade-offs exist between the desirable characteristics of local explanations.•A roadmap is proposed to choose the most appropriate explanation method.