Terentyev O. Models and methods of construction and analysis of the Bayesian networks for the intellectual data analysis

The thesis deals with the problem of improvement of speed and quality of data mining by development of a new method for construction and application of discrete Bayesian networks (BN), methods of generation of probabilistic inference and creating on their basis of the new decision support system. A review and analysis of methods for intellectual data analysis is proposed and substantiation of efficiency of BM vehicle application is considered. To construct BN structure a heuristic learning procedure is developed that is characterized by linear complexity and uses statistical data for learning. To determine a measure of interaction between the nodes the method uses values of mutual information. As a scoring measure for determining the model structure the minimum description length is used (MDL). It has also been proposed the method for exact probabilistic inference generation which uses the matrix of empirical values of joint distribution of nodes for the whole net instead of the conditional probability table. To evaluate the quality of the net constructed a new modification of Cooper-Herskovits function is proposed that is distinguished with substantially shorter computing time and relieves the restriction to the learning data size