Abstract This paper introduces the Global Sensitivity Analysis framework, commonly used in mathematical modeling to assess input uncertainty, as a novel approach to understanding variable importance in machine learning. Specifically, we consider Random Forests and we aim to extend their utility beyond mere prediction. While Random Forests are highly accurate “black–box” models, their internal mechanisms often remain obscure. Traditional variable importance measures primarily quantify the contribution of each feature to predictive performance. We propose a generative variable importance ranking based on sensitivity analysis to detect the intrinsic importance of each input feature to the underlying data–generating process. As a result, it provides crucial insight into how the response is genuinely determined by the dependence structure of its predictors. A simulation study shows how our methodology not only offers deeper insights into model uncertainty but also significantly advances in explainable machine learning by enhancing explanatory capabilities and revealing complex relationships beyond just predictive performance.
Vannucci et al. (Mon,) studied this question.