A Performance Analysis of Boosting Algorithms for the Identification of Card Fraud

A cashless society is becoming more and more of a reality because of the quick development of technology. This has made it easier for fraudsters to commit their crimes, to a certain extent. To solve this issue, fraud classification systems are used to identify fraudulent activity after it has already occurred. In this experiment, the performance of classification models for this fraud detection analysis was built and compared using five deep boosting algorithms: Adaboost, CatBoost, Gradient Boosting, XGBoost, and LightGBM (LGBM). On the IEEE-CIS Fraud Detection Dataset, which Vesta Corporation provided, this study was carried out. The models were then evaluated using various performance metrics, including the Accuracy measure, F1 Score, and Area Under the Receiver Operating Characteristic Curve (AUC). The research goal is to identify the best boosting algorithm that exhibits high classification performance and high AUC and accuracy scores. An analysis of the boosting ensemble algorithms was conducted in order to accomplish this goal. For the experiment, the dataset was preprocessed using feature selection and scaling methods. The metric comparison showed that LGBM is ranked second and XGBoost has the best classification. Since lowering the false negative rate (FNR) is the main objective of a fraud detection system, XGBoost obtains the highest fraud detection score. Overall, XGBoost achieves the highest predictive accuracy when compared to the other methods.