Increasing the statistical signification of regression dependence in the problems of processing experimental samples of small volume

The article discusses the problem of increasing the reliability and statistical significance of regression models of research objects based on small experimental samples. The possibility of expanding the sample size by generating data clusters with centroids at experimental points of the initial sample using conditional generative-adversarial neural networks was investigated. It was found that the shift of the cluster centers of the normalized generated data from the corresponding points of the experimental sample does not exceed 0.04 according to the Euclid-Mahalonobis metric. The presence of a stable multiple correlation between the vectors of input and output parameters within the generated data clusters was revealed. The possibility of increasing the reliability of estimates of regression parameters and expanding the possibility of choosing the type of regression by including the generated data in the sample was confirmed.