Experimental investigation on editing for the k-NN rule through a genetic algorithm

Abstract

An experimental investigation on selection of a reference set for the k-Nearest Neighbors (k-NN) classification method has been conducted. Genetic algorithms have been employed bringing together the strategy to preserve the decision boundary and that of selecting the most ”typical” objects as prototypes. The chromosome is directly mapped onto the reference set and the best subset is subsequently evolved. Two fitness functions have been examined. The results are contrasted with those obtained with the whole sample (before editing), Hart's and Wilson's methods. Independent subsets have been used for training and for test. Two data sets were used: two highly overlapping Gaussian classes and a data set from neonatology. The results with the proposed editing technique compare favorably with those obtained with the classical methods and with the non-edited sample.