Dyulicheva Y. Pruned Binary Decision Tree Correction Models.

The dissertation is dedicated to research and improvement of learning and recognition algorithms based on building up binary decision trees; to working out the rules for binary decision trees pruning on the basis of conjunctive regularity evaluation; to creating consistent procedure of decision tree family synthesis (i.e. the empirical decision forest synthesis algorithm) and pruned decision trees family correction methods as a set of heuristic procedures for decision making. The probabilistic decision tree (DT) pruning criterion was worked out which applies to the branches with the number of internal nodes exceeding the predetermined value of r rank. The grounding for the pruning as viewed as non-randomness of detecting r rank conjunctive regularity in the empirical sample is suggested in the work. The methods for building up a correct decision tree family so called empirical decision forest were worked out which offers a possibility for accurate fitting for a training sample, with the restriction applying to DT rank branches being observed. The appropriateness of further complication of recognition rules for building up decision making correction procedures is grounded on the basis of VCD (with the decision rules complexity according to Vapnik-Chervonenkis theory) evaluation for recognition class algorithms that are defined by the binary decision tree with the pruning applying to the number of nodes. The algebraic correction model of the incorrect empirical decision trees family was worked out which gives way to more accurate classification. The software was created to implement the algorithms introduced in the dissertation, and experiments had been carried out with real data involved that justified the theoretical results reached.