EVALUATION OF THE EFFECTIVENESS OF ENSEMBLE CLASSIFIERS AND SINGLE NEURAL NETWORK MODELS FOR FINANCIAL FRAUD DETECTION
Nikita M. Krainov, Leonid K. Bobrov
Novosibirsk State University of Economics and Management, Novosibirsk, Russian Federation
Keywords: fraud monitoring, systematic review, artificial intelligence, machine learning, models, algorithms, ranking, experimental testing
Abstract
Widely used traditional rule-based fraud monitoring systems, without ap-propriate modernization using artificial intelligence tools, are unable to effectively combat fraudulent activities, which increasingly rely on new technological advances. This article presents a systematic review of Russian literature devoted to the applicability of various machine learning models and algorithms in relation to fraud monitoring tasks. The purpose of the proposed study is to identify promising areas for modernizing traditional fraud monitoring systems based on the identification, systematization, ranking, and experimental testing of machine learning models and algorithms, as well as a comparison of the obtained results with similar studies by foreign authors, which were carried out through the analysis of the flow of foreign publications. The PRISMA protocol was used as a methodological foundation for creating the review. The initial sample included 853 publications from the Russian eLibrary and CyberLeninka databases. The final sample, after all stages of analysis using the PRISMA method, included 39 published papers. Sixteen models and algorithms were considered, including deep learning models (LSTM, GNN, CYCLEGAN), optimization algorithms (GA), hybrids of quantum computing with deep learning models, classical (Random Forest, Isolation Forest, Decision Tree) and statistical models (Clusterization, Regression, Bayes), gradient boosting (XGBoost, AdaBoost, CatBoost), and data balancing models (SMOTE, SMOTENN). An analysis of publications revealed that Random Forest (first place) and Logistic Regression (second place) are the most frequently used algorithms in building anti-fraud systems, while Isolation Forest and the Bayesian algorithm are among the least popular. The obtained results correlate with studies of the flow of English-language publications reflected in the works of foreign authors. As part of the experimental study, the first-place algorithm, Random Forest, was tested. Testing was conducted on an array of 2.2 million anonymized transactions of bank clients. As a result, the Recall metric reached 0.941, the F1-score was 0.9267, the ROC-AUC was 0.923, and the Precision was 0.9188. Accuracy was 0.9741. This is consistent with the results of similar experiments by foreign authors and provides practical confirmation of the preferable choice of the Random Forest algorithm for detecting fraudulent transactions.
|
