A Novel Kernel Extension for the Nearest Feature Line Classifier

Abstract

The Nearest Feature Line (NFL) rule represents a widely applied variant of the Nearest Neighbor rule whose usefulness has been shown in many different contexts. However, even though the NFL rule has been extended in different directions, only one is based on kernels. In this paper we make one step forward, proposing a novel kernel extension of the NFL rule called Kernel Rectified NFL classifier. Our approach is based on tools and concepts coming from the Rectified Nearest Feature Line Segment (RNFLS) classifier, another extension of the NFL rule aimed at solving known problems of the original one, such as extrapolation and interpolation inaccuracies. In the paper, we present the method and an empirical evaluation based on 15 different datasets, which show that kernelization is a promising direction for extending the NFL family of classifiers.