Machine learning, neural networks

Influence of cluster data structure on neural network model predictive properties

A. A. Khalafyan, Z. A. Temerdashev, A. G. Abakumov

Kuban State University, Krasnodar, Russia

Abstract: Theoretical works in the field of stability and convergence of neural networks include the Kolmogorov–Arnold theorem, Vapnik–Chervonenkis theory of learning, the universal approximation theorem, the combinatorial theory of precedent reliability and other. They allowed us to obtain theoretical estimates of the required volumes of training samples for building reliable models with high generalizing properties. However, the estimates turned out to be too high for practical application. Nevertheless, the results of the work served as the basis for the developers' desire to increase the amount of training data, the duration of training, and the capacity of the model in order to increase generalizing properties. Some people use a 1:10 ratio of the number of predictors of the model to the number of observations, assuming that the larger the right side of the ratio, the more reliable the model. Using a practical example, the paper shows the influence of structural features of data on the predictive properties of neural networks when solving classification problems.

Keywords: classification, cluster structure, neural network, overfitting, performance, scatterplot, canonical values.

DOI: 10.14357/20718594250202

Bibliographic databases:

•