A General Framework for Inference on Algorithm-Agnostic Variable Importance

In many applications, it is of interest to assess the relative contribution features (or subsets features) toward goal predicting a response—in other words, gauge variable importance features. Most recent work on assessment has focused describing within confines given prediction algorithm. However, such does not necessarily characterize potential features, and may provide misleading reflection intrinsic value these To address this limitation, we propose general framework for nonparametric inference interpretable algorithm-agnostic importance. We define as population-level contrast between oracle predictiveness all available versus except those under consideration. efficient estimation procedure that allows construction valid confidence intervals, even when machine learning techniques are used. also outline strategy testing null hypothesis. Through simulations, show our proposal good operating characteristics, illustrate its use with data from study an antibody against HIV-1 infection. Supplementary materials article online.