Abstract
Credit card fraud occurs most often in online purchases; therefore, it is crucial to employ better ways to find it to avoid financial loss. This paper discusses fraud detection by employing methods to generate synthetic data to improve detection models. We use the Kaggle credit card transaction dataset, implementing synthetic data generation using SMOTE as a way to balance the dataset, in which fraud cases comprise only 0.2% of cases, and perform feature engineering to better understand buying behavior. We experimented with five ML models–XGBoost, LightGBM, Random Forest, Neural Networks, and Logistic Regression; focusing on precision, recall, F1-score, and accuracy. The comparison indicates that XGBoost achieves its highest F1-score (82.57%) with good precision (93.75%) and recall (73.77%), indicating XGBoost can balance false positives and false negatives. Although all models performed with high accuracy (over 99.9%), this research focuses on highlighting precision and recall in fraud detection. The findings suggest that combining synthetic data with gradient-boosting algorithms can help fraud detection systems improve the security of online purchases.