Artificial Intelligence-Based Empirical Study for Fraud Detection in Digital Financial Transactions

Abstract

Fraud in digital financial transactions presents a critical challenge due to increasing transaction volumes and complex behavioral patterns, particularly under highly imbalanced conditions where fraudulent cases are extremely rare. This study aims to develop artificial intelligence-based models for fraud detection, compare multiple machine learning algorithms, and evaluate their effectiveness using robust performance metrics. A large-scale transaction dataset comprising over 1.32 million records with a fraud rate of approximately 0.13% was analyzed. Data preprocessing, feature engineering, and leakage-aware modelling were performed, and a chronological split (70–15–15) was adopted for training, validation, and testing. Logistic Regression, Random Forest, and XGBoost models were implemented, with threshold optimization conducted on the validation set. Performance was evaluated using accuracy, precision, recall, F1-score, ROC-AUC, and PR-AUC. The results indicate that XGBoost achieved the best performance with precision = 0.99, recall = 0.96, F1-score = 0.98, ROC-AUC = 0.99, and PR-AUC = 0.97, outperforming Random Forest and Logistic Regression. Feature importance analysis revealed that transaction-to-balance ratios, sender balance characteristics, and behavioral indicators were the most influential predictors. Overall, the findings demonstrate that ensemble-based artificial intelligence models, particularly XGBoost, provide accurate and reliable fraud detection under imbalanced conditions, supporting their applicability in real-world financial security systems.