Comparison of Multi-objective Evolutionary Algorithms for Prototype Selection in Nearest Neighbor Classification

The nearest neighbor classifiers are popular supervised classifiers due to their ease of use and good performance. However, in spite of their success, they suffer from some defects such as high storage requirements, high computational complexity, and low noise tolerance. In order to address these drawbacks, prototype selection has been studied as a technique to reduce the size of training datasets without deprecating the classification accuracy. Due to the need of achieving a trade-off between accuracy and reduction, Multi-Objective Evolutionary Algorithms (MOEAs) are emerging as methods efficient in solving the prototype selection problem. The goal of this paper is to perform a systematic comparison among well-known MOEAs in order to study their effects in solving this problem. The comparison involves the study of MOEAs' performance in terms of the well-known measures such as hypervolume, Delta index and coverage of two sets. The empirical analysis of the experimental results is validated through a statistical multiple comparison procedure.