Intelligent Feature and Instance Selection to Improve Nearest Neighbor Classifiers

Abstract

AbstractFeature and instance selection before classification is a very important task, which can lead to big improvements in both classifier accuracy and classifier speed. However, few papers consider the simultaneous or combined instance and feature selection for Nearest Neighbor classifiers in a deterministic way. This paper proposes a novel deterministic feature and instance selection algorithm, which uses the recently introduced Minimum Neighborhood Rough Sets as basis for the selection process. The algorithm relies on a metadata computation to guide instance selection. The proposed algorithm deals with mixed and incomplete data and arbitrarily dissimilarity functions. Numerical experiments over repository databases were carried out to compare the proposal with respect to previous methods and to the classifier using the original sample. These experiments show the proposal has a good performance according to classifier accuracy and instance and feature reduction.Keywordsinstance selectionobject selectionrough setsnearest neighbor