An experimental comparison of nonparametric classifiers for time-constrained classification tasks

For nonparametric and time-constrained classification problems, classifiers used are less standard, and we compare the properties of a number of such classifiers in this paper. They include: the edited/condensed nearest neighbour classifier, the reduced Parzen classifier, the neural network classifier, and all improved version of the edited/condensed nearest neighbour classifier. By means of a large series of experiments: 1) we investigate the small training sample behaviour of these classifiers; 2) we establish which classifier has the best performance for a given amount of classification time; and 3) we determine which classifier has the best overall performance. The experiments show that of the four classifiers that we have investigated neural network classifiers are best suited for time-constrained nonparametric classification problem. The results obtained have an impact on texture recognition problems, where a huge amount of pixel/voxel attributes are needed to be processed quickly