БАЛАНСУВАННЯ ДАНИХ У НЕЙРОМЕРЕЖЕВИХ СИСТЕМАХ ВИЯВЛЕННЯ ФІНАНСОВОГО ШАХРАЙСТВА
Актуальні проблеми сталого розвитку · 01.06.2026 · 3 (6) · 170-181
DOI: 10.60022/3(6)-19s ↗Kurz gefasst
Fachartikel in der Zeitschrift Актуальні проблеми сталого розвитку, online erschienen am 01.06.2026. Verfasst von Гліб Валерійович Пекуровський und 1 weiteren Autor:innen. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).
Abstract (Originalsprache)
This paper investigates the critical problem of detecting fraudulent banking transactions under severe class imbalance conditions, which is a common challenge in modern financial data science. In real-world datasets, the number of legitimate transactions significantly outnumbers fraudulent ones, severely hindering the training process and predictive accuracy of machine learning algorithms. To address this issue, this study evaluates and compares four distinct training sample balancing strategies: the baseline approach (no balancing), class weighting, the Synthetic Minority Over-sampling Technique (SMOTE), and random under-sampling. The classification task is performed using an artificial neural network configured as a Multilayer Perceptron (MLP). Due to the extreme disproportion of classes, traditional evaluation metrics such as standard accuracy fail to provide an objective assessment of model performance. To overcome this limitation, the authors propose a novel metric — Weighted Detection Efficiency (WDE). This metric normalizes the overall quality of the model by incorporating the class imbalance coefficient, thereby enabling a statistically mathematically justified and correct comparison between different balancing techniques. Experimental results demonstrated that while the baseline model without any balancing achieved the highest nominal WDE value due to its high precision on the majority class, it suffered from a critically low recall rate, failing to detect the majority of actual fraud cases. On the other hand, resampling techniques provided a more robust representation of the minority class. The synthesis of the experimental data indicates that the SMOTE method delivers the optimal balance between precision and recall, significantly improving the model’s ability to identify fraudulent activities without generating an unacceptable volume of false positives. The findings of this research can be practically applied to enhance the reliability of automated anti-fraud monitoring systems in banking and financial institutions.
Zitieren
APA
Пекуровський, Г. В., & Пекуровський, Є. В. (2026). БАЛАНСУВАННЯ ДАНИХ У НЕЙРОМЕРЕЖЕВИХ СИСТЕМАХ ВИЯВЛЕННЯ ФІНАНСОВОГО ШАХРАЙСТВА. Актуальні проблеми сталого розвитку, 3(6), 170-181. https://doi.org/10.60022/3(6)-19s
Vancouver
Пекуровський ГВ, Пекуровський ЄВ. БАЛАНСУВАННЯ ДАНИХ У НЕЙРОМЕРЕЖЕВИХ СИСТЕМАХ ВИЯВЛЕННЯ ФІНАНСОВОГО ШАХРАЙСТВА. Актуальні проблеми сталого розвитку. 2026;3(6):170-181. doi:10.60022/3(6)-19s
BibTeX
@article{2026,
author = {Пекуровський, Гліб Валерійович and Пекуровський, Єгор Валерійович},
title = {БАЛАНСУВАННЯ ДАНИХ У НЕЙРОМЕРЕЖЕВИХ СИСТЕМАХ ВИЯВЛЕННЯ ФІНАНСОВОГО ШАХРАЙСТВА},
journal = {Актуальні проблеми сталого розвитку},
year = {2026},
volume = {3},
number = {6},
pages = {170--181},
doi = {10.60022/3(6)-19s}
}
EUVIMED