Povkhan I. Methods and principles of constructing discrete object classification trees for data mining.

Present thesis is devoted to studying the specific features of both logical and algorithmic classification trees, some issues of their use in the problems of intelligence data analysis and image recognition by creating the efficient models of classification and analysis of large and super-large data arrays and universal systems of discrete object recognition. Tree-like classification models of various types are analyzed and studied. The concept of constructing tree-like classification models based on a step-by-step approximation of an array of initial TS data with a set of selected and evaluated independent recognition algorithms is proposed. The concept of T – reference sets is proposed, which consists in selecting and fixing a set of features together with their classifier values in the LCT structure. Models and methods for constructing ACT structures of two types are developed, where the resulting classification trees consist of various algorithms and recognition methods, which in turn represent new algorithms and classification schemes. A model and method for constructing bounded ACT structures are developed, which are aimed at completing only those paths, tiers of the classification tree structure where there are the largest number of errors of all classification types. Methods for finding the similarity of logical tree structures in problems of minimizing their structures are developed. A numerical estimate of the impact of the pruning procedure of the constructed LCT – rearrangement of tiers, levels, and blocks of the LCT structure in the construction of a regular logical tree on its overall complexity is given. The convergence of the procedure for constructing classification tree models proposed in the study of ACT structures for conditions of weak and strong separation of initial TS classes is estimated.