Performance Comparison of Prototype Selection Based on Edition Search for Nearest Neighbor Classification
Nordiana Binti Mukahar, Bakhtiar Affendi Rosdi · 2018
The k-nearest neighbor classifier is one of the most commonly used non-parametric classifiers in pattern recognition fields. It received wide interest among researchers due to its effectiveness in its implementation and simplicity. However, the k-nearest neighbor classifier suffers from several deficiencies such as high storage requirements, low computational efficiencies and low degree of noise tolerance. Several prototype selection techniques have been proposed in the literature to overcome the drawbacks in the k-nearest neighbor classifier. Prototype selection involves preprocessing the training set with editing algorithm to remove irrelevant samples. This paper conducts an experimental comparison of various prototype selection techniques based on edition search in nearest neighbor classification on ten real UCI data sets. Empirically, the experimental study involving different sizes of real UCI data sets and measuring the performance in terms of the classification accuracy, reduction capabilities and computational time.