ПРОГНОЗУВАННЯ ЧАСОВИХ РЯДІВ СОЦІАЛЬНИХ МЕРЕЖ НА ОСНОВІ НЕЙРОМЕРЕЖЕВОГО ПІДХОДУ

Abstract

The article presents a predictive model for analyzing the state and behavior of social network objects in dynamics and for building predictions of user behavior. This opens up opportunities for choosing the most promising promotion of objects in social networks, the best placement of advertising, the possibility of creating independent projects. The article proposes an approach to the formalization of the processes of growth/decrease in the popularity of content under the influence of external factors within the framework of a mathematical predictive model based on a recurrent neural network. The purpose of the model is mathematical processing and analysis of stochastic information flows, graphical presentation of forecasts of music trends, changes in the popularity of bloggers, video content, and more. The developed model can be used to predict the growth of reposts in social networks. The list of external and internal variables that affect the behavior of the output data at each moment of time is described. The behavior of time series is studied on the basis of input data obtained from social networks (nomenclature of indicators, number of reposts of researched objects, simulation period). A feature of the model is the ability to work with Big Data arrays, which in most cases have incomplete or "noisy" information. The dependence between input and output is proposed to be specified using a system of constraints and external coefficients that change over time. Control parameters are predictors of the model. Features of building a predictive model are shown, a list of the main limitations of the mathematical model is described. The result of the implementation of the tasks set in the research is a dynamic predictive model that provides an opportunity to create predictions based on such indicators as reposts on social networks, music trends, and the popularity of bloggers. It is proposed to use a recurrent artificial network with a dynamic neuron to predict reposts on selected research objects. The neuron processes input signals and calculates a predictive indicator of the number of reposts. The learning process of the neural network during the simulation period of one month is demonstrated. To reveal the prospects of further development of the model in order to improve the accuracy of forecasts of socio-economic processes.