Methods to compare nonparametric classifiers and to select the predictors

Abstract

In this paper we examine some nonparametric evaluation methods to compare the prediction capability of supervised classification models. We show also the importance, in nonparametric models, to eliminate the noise variables with a simple selection procedure. It is shown that a simpler model usually gives lower prediction error and is more interpretable. We show some empirical results applying nonparametric classification models on real and artificial data sets.