Voting Ensemble Framework with SMOTE for Financial Fraud Detection in Highly Imbalanced Transaction Data (INTERANTIONAL JOURNAL OF SCIENTIFIC RESEARCH IN ENGINEERING AND MANAGEMENT)

Summary

• The framework addresses extreme class imbalance in a dataset comprising 6,362,620 transaction records, where only 8,213 records (0.1291%) represent fraudulent activity.p.1
• Data preprocessing transformed the initial dataset shape from 6,362,620 × 11 into a engineered matrix of 6,362,620 × 16 attributes.p.3
• The experimental evaluation used an unmanipulated test subset of 1,272,524 transactions containing 1,270,881 legitimate records and 1,643 fraudulent records.p.4
• While the Voting Ensemble achieved the top F1-score of 0.98204, the Stacking Ensemble + SMOTE configuration attained the highest recall score of 0.99878.p.6

Key insights

1. Traditional threshold-based fraud monitoring systems lack the capability to detect emerging fraud patterns compared to modern machine learning algorithms.p.1
2. Account identifiers nameOrig and nameDest were intentionally removed prior to training to eliminate data leakage and account-specific memorization.p.3
3. Continuous transaction variables were standardized using parameters calculated strictly from training data to maintain evaluation integrity on unseen test samples.p.4
4. Ensemble and boosting methods consistently outperformed neural networks and transformer-based architectures across all primary classification metrics.p.5
5. Stacking Ensemble + SMOTE obtained the highest recall rate of 0.99878, though with slightly lower precision than the soft voting ensemble.p.6
6. Deep learning architectures combined with Focal Loss, including 1D CNN and Transformer models, recorded lower overall F1-scores than tree-based ensemble methods.p.6