UDC 004.032.26
DOI: 10.36871/2618-9976.2022.11.007
Authors
Victor V. Erokhin
Doctor of Technical Science, Docent, Moscow state institute of international relations (university),
Ministry of Foreign Affairs of the Russian Federation, Moscow, Russia
Elena V. Eliseeva
Candidate of Pedagogical Sciences, Docent, Bryansk State Academician I.G. Petrovski University,
Bryansk, Russia
Abstract
An integrated structure from a discrete choice model with an embedded artificial neural network is proposed: TN-MNLR, which allows more flexible and accurate solving of regression problems with good predictability. The TN-MNLR framework allows you to more accurately evaluate Multinomial logit regression (MNLR) built on nonlinear input parameters. An artificial neural network with feedforward is used to predict the parameters of the studied regression in the form of a nonlinear function. Further, these found parameters from the neural network are transferred to the parametric logical data selection model to calculate the probabilities of their selection. Input parameter constraints are applied in the neural network transformation layers. The strengths and weaknesses of using the TN-MNLR structure for econometric discrete data selection models are presented, which makes it possible to improve the accuracy of big data estimation during their multivariate optimization. The goal has been achieved to ensure the scalability and behavioral interpretability of the results of the structure used, which also contributes to the acquisition of new knowledge from input discrete big data with their detailing. However, the application of the presented structure is limited only for heterogeneous data. The structures under consideration can effectively solve transport problems and the tasks of managing organizational systems.
Keywords
Artificial neural networks, Neural network hyperparameters, Deep neural networks, Modeling, Discrete choice models, Multinomial logit regression
