2015 , Volume 20, № 6, p.36-45

Purpose. Search for solutions of the following problems: - to find the minimum cover of a training set E₀={S₁,…,S_m} by standard objects. The set E₀ is divided into l(l≥2) of disjoint subsets (classes) K₁,…,K_l. The objects are described by a set of features X(n)=(x₁,…,x_n);- a reduction of the dimension of the features space by constructing groups from the disjoint sets of features X(k₁ ),…,X(k_p ),k₁+⋯+k_p≤n and nonlinear mapping of their values on the real axis in the description of objects. Methodology. The method of partitioning the training set E₀ into disjoint groups of objects using the property of their connectivity through the subset of boundary objects (shell) of classes L(E₀,ρ) on a metric ρ(x,y) was developed. The set G_tν=S_ν1 ,…,S_νc, c≥2, G_tν⊂K_t, v<|K_t| presents an area (group) with the constrained objects in the class K_t, if for any S_νi ,S_νj∈G_tν the pathway S_νi↔S_νk↔⋯↔S_νj will exist. The standard objects of minimum covering of training set E₀ are selected for each group G_tν separately. The rule of hierarchical clustering of features offers for the nonlinear mapping of their values on the real axis. The use of this rule allows: - to form a new space from latent features; - to produce the ordered selection of informative features. Findings. It is proved, that the number of standard objects for covering which ensures the correct recognition on precedents in the training set, does not increase monotonously when the number of features with low value of information are reduced. Originality/value. Complex use of methods of cluster analysis for both group objects and features allows to reduce the volume of training sets and to increase the stability of the internally defined logical regularities.