main| new issue| archive| editorial board| for the authors| publishing house|
Main page
New issue
Archive of articles
Editorial board
For the authors
Publishing house



No. 9. Vol. 25. 2019

DOI: 10.17587/it.25.545-552

N. A. Ignatiev, D. Sc., Professor, e-mail:, National University of Uzbekistan named after M. Ulugbek, Tashkent

Compactness of Objects of Classes and Selection of Learning Samples

The relations between the sample objects E0 = {S1,…, Sm}, divided into disjoint classes K1, Kl, is considered. Compactness determines the quantitative measure of these relations by a given metric r(x, y) and various subsets of the set of features X(n). The selection of the sample is based on the results of the sequential execution of three operators:

  • search and delete noise objects;
  • minimum sample coverage with objects—standards;
  • selection informative set of features.

The rule according to which noise objects are considered as a subset of boundary objects of classes is described. A grouping algorithm based on the relationship "'<->" of connectedness objects of classes has been developed. For any pair (Si, Sj), Si, Sj ÎKd , d = 1,…, l, included in one group, there always exists a chain Si" Sk" ..." Sj. The search for the minimum sample coverage with reference objects is performed for each group separately. It is shown that when selecting informative features, the average number of objects attracted by one of the standard increases.
Keywords: noise object, informative features, selection of training samples, object—standard, generalizing ability, cross validation



To the contents