On the reduction of the nearest-neighbor variation for more accurate classification and error estimates

Abstract

In designing the nearest-neighbor (NN) classifier, a method is presented to produce a finite sample size risk close to the asymptotic one. It is based on an attempt to eliminate the first-order effects of the sample size, as well as all higher odd terms. This method uses the 2-NN rule without the rejection option and utilizes a polarization scheme. Simulation results are included as a means of verifying this analysis.