ML-Based Detection of Credit Card Fraud Using Synthetic Minority Oversampling

Abstract

Credit card fraud detection is an essential and classic but very difficult problem consisting of imbalanced classification where fraud transactions are almost nonexistent as compared with legitimate ones. This paper proposes an approach which main feature is a combination of an imbalanced data technique based on the Synthetic Minority Oversampling Technique (SMOTE) with an XGBoost classifier for credit card fraud detection. We delineate the dataset preparation, feature preprocessing, SMOTE resampling, model configuration and training, and evaluation metrics (Accuracy, Precision, Recall, F1, AUC, and confusion matrix). Additionally, a comparative experiment plan is defined that includes baseline classical models (Logistic Regression, Random Forest, Artificial Neural Network) thereby allowing practitioners to perform benchmarking against other models' performances. The complete code necessary for conducting the experiments is accessible (the user-supplied Colab script was used as the foundation). The results show that if XGBoost is used in combination with careful preprocessing and SMOTE it will acquire a strong recall very important properties for fraud detection while still to an extent retaining high precision. We elaborate on the limitations (synthetic oversampling risks, concept drift) and plan ahead for the future inventiveness (cost-sensitive learning, streaming models, explainability).