Manikaeva O. Classification methods of heterogeneous data in intellectual systems of medical and social monitoring

The thesis is devoted to the solution of the scientific and practical problem of creation of intelligent systems (IS) of medical and sociological monitoring (MSM) on the basis of the improved models of representation of heterogeneous MSM’s data and the developed methods of their classification by means of Kohonen network training. In the research improve the models of representation of weakly structured and weakly labeled heterogeneous MSM’s data in the spaces of properties and features, taking into account values, types, formats, sources, quality assessments and procedures of aggregation / transformation of properties of detailed data. Based on these models, a method of matching classes and clusters markers in learning of Kohonen network with partial teacher involvement was developed, which is based on constructing a two-dimensional histogram of pairwise matches of classes and clusters markers values with its subsequent intersecting by rows and by columns to the developed rule. The method allows to obtain additional class markers in the unlabeled part of the training sample. Based on the method of matching classes and clusters markers to assess the suitability of chromosomes in the population of the genetic algorithm for each example from the training sample, a method of heuristic weight adjustment in the learning process of the Kohonen neural network is proposed. The use of such an adjustment of the weights allowed to reduce the training time of the Kohonen network without losing the level of reliability of the classification. The method of classification of weakly structured and weakly labeled heterogeneous MSM’s data has been improved due to the use of developed methods of matching class and cluster markers and heuristic adjustment in the process of learning the Kohonen network.