Abstract
The objective of this study is to introduce a new model of data classification based on preliminary reduction of the training set of examples (preprocessing) in order to facilitate the use of nearest neighbours (NN) techniques in near real-time applications. This study accordingly addresses the issue of minimising the computational resource requirements of NN techniques, memory as well as time. The approach proposed in the paper is a modification of the classical k-Nearest Neighbours (k-NN) method and the k-NN method with local metric induction. Generally, the k-NN method with local metric induction in comparison with the classical k-NN method gives better results in the classification of new examples. Nevertheless, for the large data sets the k-NN method with local metric induction is less time effective than the classical one. The time/space efficiency of classifying algorithms based on these two methods depends not only on a given metric but also on the size of training data. In the paper, we present three methods of preliminary reduction of the training set of examples. All reduction methods decrease the size of a given experimental data preserving the relatively high classification accuracy. Results of experiments conducted on well known data sets, demonstrate the potential benefits of such reduction methods.