A warning signal for variable importance interpretation in tree-based algorithms

Tree-based learning algorithms are largely utilized in several scientific domains. Each algorithm specifies a proper variable importance measure that captures the magnitude of the predictive importance of an explanatory variable. In applied studies, such measures are sometimes interpreted as importance relative to the data generating process, but this can be misleading. To prevent erroneous interpretation of the variable importance there is a need of a measure capable of detecting when the predictive variable importance differs from the explanatory one. We propose a measure that can act as a warning signal of this situation.