An Ensemble Method for Feature Screening

It is known that feature selection/screening for high-dimensional nonparametric models an important but very difficult issue. In this paper, we first point out the limitations of existing screening methods. particular, model-free sure independence methods, which are defined on random predictors, may completely miss some features in underlying function when predictors follow certain distributions. To overcome these limitations, propose ensemble procedure models. elaborately combines several methods and outputs a result close to best one Numerical examples indicate proposed method competitive has satisfactory performance even fail.