CLASS OVERLAP RATE AS A DESIGN CRITERION FOR A PARALLEL NEAREST NEIGHBOUR CLASSIFIER
Leszek J. Chmielewski, Marek Skłodowski, Waldemar Cudny · 1999
It is well known that the classifiers based on the 1-NN and the k-NN rules offer good performance and that their implementation is simple. However, classification of a single object requires calculation of a large number of distances, equal to the power of the reference set. The classification speed depends also on the number of features of the objects. This paper deals with the problems of feature selection and reference set reduction as the methods of improving the classification speed. As a classifier the parallel net of two-decision k-NN classifiers is considered. Each of the component k-NN classifiers is approximated by a 1-NN classifier with the reduced reference set. The proposed classifier is experimentally compared with the standard k-NN one. a very requiring data set obtained in an industrial application to quality inspection of the surfaces of ferrite products has been used in the experiments.