EVALUATION OF THE EFFECTIVENESS OF ENSEMBLE CLASSIFIERS AND SINGLE NEURAL NETWORK MODELS FOR FINANCIAL FRAUD DETECTION
Andrey I. Pestunov, Roman V. Samoylenko
Novosibirsk State University of Economics and Management, Novosibirsk, Russian Federation
Keywords: recurrent neural network, convolutional neural network, ensemble classifier, financial fraud detection
Abstract
This study investigates the effectiveness of single and ensemble classifiers for financial fraud detection based on account-to-account money transfer data. We considered Recurrent Neural Networks, Convolutional Neural Networks, Autoencoders, as well as ensemble methods such as Random Forest, XGBoost, and stacking; simple voting was additionally applied to the neural network models. Experiments were conducted using a large-scale, real-world labeled dataset of transaction records. Classification performance was evaluated using Precision, Recall, F1-score, and ROC-AUC metrics. Computational efficiency was assessed based on training time, inference time, memory usage, and average CPU load. The experimental results demonstrated that the stacking ensemble method achieves the highest accuracy, although it is the most resource-intensive. Meanwhile, the accuracy of more efficient boosting and bagging methods is slightly lower but remains comparable to stacking.
|
