Selecting classifiers to ensure the quality and reliability of pattern recognition at class intersection

Abstract

The article poses and solves the problem of selecting classifiers for the case when the given classes intersect in the space of initial properties of objects. The value of the limiting dimension of the classifier space is determined, taking into account the volume of the learning sample, the probability of errors made when separating classes and recognizing new objects. Procedures are given for determining the classifiers of three types and selecting among them the ones that satisfy the predetermined error when separating classes, and the predetermined error probability and reliability when recognizing new objects. An algorithm and software were developed based on the proposed procedures. Computational experiments on a computer and the results were presented in the form of decision rules used for object recognition.