STACKING ENSEMBLE, XGBOOST DAN SMOTE UNTUK EFISIENSI ENERGI PADA FRAUD CREDIT CARD

Extreme class imbalance in credit card fraud detection datasets often leads machine learning models to fail in recognizing minority fraud cases. This study proposes a Stacking Ensemble approach combined with Synthetic Minority Oversampling Technique (SMOTE) and SMOTE-Tomek Links, employing XGBoost, LightGBM, and Random Forest as base learners and XGBoost as a meta-learner. Using the Kaggle Credit Card Fraud Detection dataset (class ratio 1:492), the method was evaluated with Recall, F1-Score, AUC-PR, and AUC-ROC, while CodeCarbon was integrated to measure energy consumption and carbon emissions during model training. Experimental results show that the proposed ensemble improves Recall of fraud detection by up to 6% compared to single models, achieves stable F1-Scores of 0.92 (SMOTE) and 0.91 (SMOTE-Tomek), and records an AUC-PR above 0.90. Furthermore, CodeCarbon tracking indicates that SMOTE models produce slightly lower carbon emissions (0.62 gCO₂) than SMOTE-Tomek (0.68 gCO₂), highlighting a trade-off between detection accuracy and energy efficiency. These findings emphasize that integrating ensemble learning with oversampling techniques not only enhances fraud detection performance but also provides transparent insights into the environmental impact of machine learning models.