Multivariate data visualization for qualitative model choice in learning systems

An visualization of multidimensional nonlinear structures of multivariate data for a qualitative model choice are considered. These problems appear in technological, pharmaceutical, financial, medical, etc., industries. It is showed that conventional methods of data analysis and model choice could be essentially reinforced by means of more effective use of human vision abilities. The pictorial analysis is developed as a human-computer dialog to search "pictorial" data representations which allow an effective detection of informative features and dependencies. Basic problems of this approach are discussed: the choice of perspective class of pictorial representation for different data sets, the interactive search of representations which ones allow to detect and describe visually acceptable solution of problem (pattern recognition, clustering), the estimation of possible solution time and complexity for different data structures, etc. Practical implementations showing a high effectiveness of pictorial methods are demonstrated in application to a gas-turbine engine diagnostics by nonstationary parameter trends, for an simultaneous monitoring of a hundred parameter dynamics of a power plant and for search of structure-activity relationship for new organic compounds.