Structure Choice for Relations between Objects in Metric Classification Algorithms

Abstract

We analyze the cluster structure of learning samples, decomposing class objects into disjoint groups. Decomposition results are used for the computation of the compactness measure for the sample and its minimal coverage by standard objects. We show that the number of standard objects depends on the metric choice, the distance to noise objects, the scales of the feature measurements, and nonlinear transformations of the feature space. We experimentally prove that the set of standards of the minimal coverage and noise objects affect the algorithm generalizing ability.