A sequential approach to feature selection in high-dimensional additive models

Abstract

We deal with the problem of feature selection for high-dimensional additive models in this article. The existing feature selection methods for additive models in the literature mainly concentrate on the penalized likelihood approach. We propose a sequential group selection method for additive models (sgsAM) in this article. The additive functions are estimated by B-spline method and are treated as groups of features. The groups are selected sequentially by a correlation search procedure. It is shown that the sgsAM method is selection consistent. Numerical studies using simulated data and a real data demonstrate that the sgsAM method has an advantage over the existing methods.