Neural network method for automatic data generation in adaptive information systems

The paper discusses the development of a new method for automatic generation of information in adaptive information systems (AIS), based on the application of neural network technologies. The analysis of existing approaches for the solution of this problem is carried out, on the basis of which a conclusion is made about the prospects of using neural networks for information generation in AIS. Prevailing public methods (AutoKeras, DEvol, AutoSklearn, AutoGAN) do not allow solving the whole range of selected problems of generating information in AIS in automatic mode. The analysis showed that these methods do not satisfy the following conditions: support for multidimensional input and output data, formation of independent outputs, work with generative adversarial networks and autoencoders, quality assessment of generated output objects of arbitrary type. Therefore, a neural network method for automatic generation of information in AIS was developed; its formalization in the notation of set theory, theoretical grounding, algorithms for solving specific problems in accordance with the developed method are presented. During the experimental research, the practical implementation of the neural network method and its comparison with existing analogues were conducted. The scientific novelty of the method consists in the automatic determination of the class of data generation problems and operations for the preparation and processing of initial data, the use of an iterative approach for the automatic selection of the structure and parameters of the neural network, adaptation of Fréchet Inception Distance and Inception Score metrics for evaluating arbitrary data. The practical significance of the method is shown in the versatility of the approach, increasing the accuracy and reducing the search time for the optimal structure of neural networks, the possibility of integrating different architectures of neural networks (including generative adversarial and autoencoders) for automatic solution of information generation problems. Comparing the neural network method with existing approaches (AutoKeras, DEvol, AutoSklearn, AutoGAN), it was found that the developed method is more functional, exhibits more speed of the structure of neural networks search (3.2 times), and provides comparable accuracy (by 2.3%).