Credit Card Fraud Detection Based on Multiple Machine Learning Models
Muyuan Chen · 2022
Credit cards have long been one of the most popular methods of making payments. But this kind of payment still carry risks —— credit card fraud. Machine learning has played an essential role in detecting credit card fraud. This study uses machine learning methods to provide most accurate prediction of fraudulent transactions. These algorithms include Random Forest, Logistic Regression, SVM, and Extreme Gradient Boosting (XGBoost). Dataset used in this study is from Kaggle and is highly skewed. In order to find out if imbalanced dataset will affect predictions of models, three oversampling methods, Random oversampling, SMOTE, and ADASYN, are used to resample dataset for comparative analysis. From the analyzed results, oversampling methods tend to cause overfitting of models. XGBoost is the only technique that are not affected by oversampling methods. Since overfitting is a problem that should be avoided in classification problem, original data is selected for further evaluation of models. Performance of four algorithms are evaluated based on F-measure, Cohen's Kappa, AUC score, and accuracy. The results show that Random Forest and XGBoost outperform SVM and Logistic Regression. As a result, a fusion model is constructed based on these two classifiers. Log loss and brier score are used to further evaluate their predictions. The results show that Random Forest and XGBoost provide best predictions and close to each other. Fusion model has a little improvement compared to these two classifiers.