The methods of modeling and structure estimation building for KNN classifiers on the basis of small training sets

Abstract

In this paper the fullest conception of the probabilistically combinatorial approach has been presented. This conception is the result of previous long preliminary works. The approach gives the possibility to establish the reasons of algorithms overtraining, to define the possible ways of it reduction and to build the most precise estimates of the recognition probability. The combinatorial approach works with determined data of the recognition process and the probabilistic one determines the probability of these results existence. The most usefulness of the combinatorial approach consists in the possibility to determine the effect of the training set variation on the different algorithms and select the most appropriate one from these algorithms or algorithm composition. The probabilistic part of this approach determines the probability of results, obtained on the basis of combinatorial approach.