International Journal of Science and Technology (IJST)

Enhanced Credit Card Fraud Detection: Behavioral Features, SMOTE-ENN Balancing, and LSTM-AdaBoost Ensembles

Credit card fraud causes substantial financial losses to both consumers and financial institutions globally because of the increasing volume of online transactions and the sophistication of fraud schemes necessitate advanced detection systems. Detection accuracy is enhanced in this research through the fusion of advanced feature engineering, behavioral analysis, and neural network ensemble methods. Preprocessing of data is done to standardize the structures, kernel behavioral features of spending patterns, locations of transaction, frequency are elicited, and class imbalance is mitigated by means of SMOTE-ENN, which is later evaluated by LSTM networks with the ADA Boost to create a robust ensemble model for credit card fraud detection with and without behavioral feature integration. The results demonstrated that ensemble learning methods, particularly AdaBoost with behavioral features, achieved the highest overall performance, yielding an F1-score of 0.9642, ROC-AUC of 0.9867, and strong precision and recall values are 0.9803, 0.9485, respectively. In contrast, LSTM models benefited significantly from behavioral features, improving their ROC-AUC from 0.7732 to 0.8632 and F1-score from 0.738 to 0.79. Interestingly, AdaBoost without SMOTE-ENN underperformed (F1 = 0.4923, ROC-AUC = 0.3512), highlighting the importance of handling class imbalance effectively. Overall, our results suggest that combining ensemble methods with behavioral insights and appropriate data balancing techniques leads to highly accurate and reliable fraud detection.