Competence Enhancement for Nearest Neighbor Classification Rule by Ranking-Based Instance Selection

Abstract

This paper introduces a novel prototype selection scheme that decides which instances to preserve using an approach that defines an order to the instances in the data sets. The order of each instance is defined by its relevance to the data set considering the similarity to their nearest eighboors. Scores are assigned to the instances. Instances surrounded by others of the same class have highest scores and have priority in the selection. Experiments performed over several classification problems show that the proposed method reduces the storage requirements and keeps or improves the classification accuracy.