Improving Credit Card Fraud Detection Performance Through Boosting and Stacking Techniques

Adiena, Campin Waladsae (2025) Improving Credit Card Fraud Detection Performance Through Boosting and Stacking Techniques. Other thesis, Institut Teknologi Sepuluh Nopember.

Text 5026211050-Undergraduate_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (2MB) | Request a copy

Abstract

In 2024, over 327,000 credit card fraud cases were reported in the United States, emphasizing the urgent need for reliable fraud detection systems. Traditional methods struggle with real-world financial data due to class imbalance, overlapping patterns, and high false positive rates—since fraudulent transactions are rare compared to legitimate ones. This final project proposes an ensemble learning framework combining Boosting and Stacking techniques to address these challenges. Boosting, implemented via XGBoost, iteratively corrects misclassifications and achieved the best performance. Stacking integrates multiple base learners—Logistic Regression, Decision Tree, Random Forest, and k-Nearest Neighbors—using either XGBoost or Logistic Regression as a meta-learner to capture diverse decision patterns. To address class imbalance, the Synthetic Minority Oversampling Technique (SMOTE) was applied to the training data, generating synthetic samples to balance the dataset. Models were evaluated using accuracy, precision, recall, and F1-score. The Boosting model achieved the highest performance with 98.80% accuracy, 0.29 precision, 0.89 recall, and 0.44 F1-score, indicating strong fraud detection while keeping false positives relatively low. In contrast, the Stacking models underperformed with F1-scores of 0.27 (XGBoost meta-learner) and 0.24 (Logistic Regression meta-learner). Baseline models like Logistic Regression and Decision Tree showed high recall but low precision, demonstrating the limitations of relying solely on accuracy in imbalanced datasets. This final project acknowledges several limitations, including reliance on synthetic data, limited model diversity, static evaluation, and restricted computational resources. Future work should explore real world datasets, additional resampling methods like ADASYN, alternative classifiers such as SVM or neural networks, and advanced strategies like threshold tuning or cost-sensitive learning. Running experiments in a high-performance computing environment is also recommended. This final project contributes to the development of more effective AI-based fraud detection systems for secure digital payments.

Item Type:	Thesis (Other)
Uncontrolled Keywords:	Credit Card, Fraud Detection, Boosting, Stacking, Ensemble Learning
Subjects:	Q Science > Q Science (General) > Q325.5 Machine learning. Support vector machines. T Technology > T Technology (General) > T57.5 Data Processing
Divisions:	Faculty of Intelligent Electrical and Informatics Technology (ELECTICS) > Information System > 57201-(S1) Undergraduate Thesis
Depositing User:	Campin Waladsae Adiena
Date Deposited:	26 Jul 2025 07:40
Last Modified:	26 Jul 2025 07:40
URI:	http://repository.its.ac.id/id/eprint/121288

Actions (login required)

View Item