and Computational Methods

doi: 10.18698/2309-3684-2022-3-8497

This article proposes a method for optimizing the dynamic budget allocation policy for an advertising campaign placed through an advertising tool built into the search engine. This method takes into account the unique features of social media marketing, provides an optimal budget allocation policy over time for one advertising campaign and minimizes the duration of the campaign, taking into account the specific budget and the desired level of coverage of each marketing segment. The model includes a general "efficiency function" that determines the relationship between the cost of an advertising bid at a given time and the number of new users shown at that time. This goal is achieved by implementing an algorithm for optimal solution of the problem of dynamic distribution of the advertising budget under certain boundary conditions, as well as by analyzing data on advertising campaign for June 2018. In the course of the study, an algorithm for optimal solution of the problem of dynamic distribution of the advertising budget under appropriate boundary conditions was implemented, examples of specific cases of the efficiency function were given and some models of real advertising campaigns of the enterprise were analyzed. Then, the data registered by the advertising agency of a particular enterprise was analyzed in relation to an advertising campaign registeredusing a built-in search engine tool for calculating bids and audience coverage for 30 days.

Чибисова А.В., Шинаков Д.С. Математическое моделирование рекламной кампании. Математическое моделирование и численные методы, 2022, № 3, с. 84–97.

doi: 10.18698/2309-3684-2022-4-8192

The mathematical derivation of the presented neural network model is demonstrated. Reduction of the classification problem to an optimization problem. Produced recon-naissance data analysis, as well as their preprocessing for further use in training classification algorithms. The architectures of neural networks were designed depending on the activation function, the number of hidden layers of the neural network and the number of neurons in the hidden layers. More than ten neural networks were trained to solve the task of credit scoring. The calculation of the learning time of neural networks was made. The solution of the problem using classical machine learning algorithms is presented. It could be seen that the standard deviation of accuracy and ROC AUC for neural networks is greater than that of a random forest. This is due to the fact that we choose the initial weights randomly and calculate the gradients not on the entire sample, but on small parts, which adds some learning error. But these deviations were not only for the worse. In the best situations, according to both metrics, neural networks showed the worst result by a couple of percent. The analysis of results is made. Comparative analysis shows that neural networks have better classification quality than classical machine learning algorithms, and also that neural networks have less training time than classical machine learning algorithms. Graphs and tables displaying the results obtained are presented.

Кадиев А.Д., Чибисова А.В. Нейросетевые методы решения задачи кредитного скоринга. Математическое моделирование и численные методы, 2022, № 4, с. 81–92.