The cognitive visualization of classifying rules extracted from data based on binary solver matrix model
Viktor Pimenov,Mikhail Voronov,Iliya Pimenov
DOI: https://doi.org/10.31799/1684-8853-2019-6-2-11
2020-01-16
Information and Control Systems
Abstract:Introduction: Huge volumes of data are generated in cyberspace or from internal information of various organizations. In order to obtain a set of output data with a clear structure, divide it into significant parts and develop rules of classification, machine learning methods are used. Most inductive methods simulate intermediate and high-level abstract categories in multidimensional space which are difficult to interpret. Purpose: Developing a model of machine learning in the form of a “white box” which explains the chosen solution using conventional production rules, along with cognitive visualizers for characterizing classes of objects. Methods: Formation of a binary decision matrix containing information about a combination of the selected informative sign values which imply the specified classes. Results: A binary decision matrix is formed automatically according to the results of cluster and discriminant analyzes. The learning procedure is reduced to setting interval thresholds and matrix elements, which makes it easy to implement a semantic interpretation of a solving rule. The object is recognized by elementwise conjunction of the matrix cells to which the values of the attributes are pointing, and by selection of a single cell corresponding to the class code. To interpret a rule, a universal algorithm for processing a binary matrix has been developed, which applies user-entered attribute values. The dimension of the viewed space is specified by adjustment rings on the recognition visualizer. The azimuth of an initiated diagram cell with the greatest dimensionality indicates the belonging of an object with the set features to a target class. For the characterization of classes, visualizers have been developed, demonstrating both the distinctive properties of a class and properties that several classes share. In many cases, the object type recognition stops when the depth of the scanned features space is significantly less than with a full search. Practical relevance: The proposed methods of cognitive analysis and data visualization provide not only the classification of data, determination of the significance of features, their ranking and selection, but also the development of rules which reveal the cause-and-effect relationship between the combination of factors and the type of a made decision.