<![CDATA[Credit card fraud remains a major challenge in the financial sector due to its dynamic nature and highly imbalanced transaction data. This study presents a robust deep ensemble learning approach that integrates spatial, sequential, and temporal learning capabilities. A series of preprocessing steps were applied, including feature normalization, class-label separation, and class rebalancing using SMOTE. The model architecture combines convolutional, recurrent, and long short-term memory layers to capture diverse fraud patterns. These components are merged and passed through dense and dropout layers for optimal binary classification. The datasets used are generated from real-world credit card transactions, ensuring practical relevance. On the test set, the proposed model achieved 99.7% accuracy, 99.6% precision, 99.9% recall, and 99.8% F1-score. The training and validation loss curves showed smooth convergence without any overfitting, confirming model stability. To ensure reliability, 3-fold stratified cross-validation was performed on the balanced dataset. The average metrics across folds included 99.76% accuracy, 99.70% precision, 99.85% recall, and 99.77% F1-score. These results underscore the generalization capability and consistent prediction performance of the model. Comparative analysis showed that the group model outperformed individual CNN, RNN, and LSTM architectures. The hybrid strategy benefits from the spatial extraction of CNN, sequence modeling of RNN, and memory retention of LSTM. By integrating these strengths, the model effectively detects subtle and complex fraud patterns. This approach provides a scalable and reliable solution for real-time fraud detection in imbalanced credit card datasets.]]>
Copyrights © 2025
