Properties of the Ordered Feature Values as a Classifier Basis

Abstract

The paper proposes a new classifier based on new concept closeness for objects finite set: feature values of the same class objects are close if the difference between these values is small enough. To pass to this concept, the combined sample data for each feature k were approximated by mapping onto a set of the ordered pairs (k;m), where m is the interval number of the feature ordered values. The objects of each pair have close values of the considered feature. Number lists of training sample objects of the same class, forming ordered pairs, was called an information granule. The frequency of any granule is calculated from the length relation of corresponding subsets as a complex event. These frequencies allow us to calculate the frequencies of the object feature values in different classes, and then the object frequencies as a whole in a certain class, the maximum of which determines the object class. Simplicity, robustness and efficiency of the developed algorithm were confirmed experimentally on 9 databases.