FeatureImp uses the median to aggregate the results of multiple permutations, which seems to be rather unusual (although I'm aware of the benefits of using a robust estimator). As far as I can tell, averaging is used in both the statistical literature, e.g. Strobl et al., and other implementations, e.g. the ingredients::feature_importance() function or the cforest, if I interpret the code correctly....
Thanks for considering this possible change or enhancement, and thanks for contributing this great package!
FeatureImp uses the median to aggregate the results of multiple permutations, which seems to be rather unusual (although I'm aware of the benefits of using a robust estimator). As far as I can tell, averaging is used in both the statistical literature, e.g. Strobl et al., and other implementations, e.g. the ingredients::feature_importance() function or the cforest, if I interpret the code correctly.... Thanks for considering this possible change or enhancement, and thanks for contributing this great package!