Abstract:
The article considers the procedure for the formation of sets of informative features, the se-lection of which is carried out according to pairs of disjoint classes of objects with the algorithm of a hierarchical agglomerative grouping of initial features. It is supposed that the class indices belong to the set of admissible values of the classification attribute in the ordinal scale. To reduce the dimen-sionality of space in the description of objects, the calculation of latent indicators by groups is used. A measure of the compactness of classes by the latent indicator is proposed, which makes it possible to evaluate sets of features. The set of features with the maximum value of the compactness measure is selected as informative. The properties of binary relations between classes with respect to the values of the compactness measure are investigated.
Keywords:nonlinear transformations, membership functions, hierarchical agglomerative grouping, generalized estimates of objects, measure of compactness.