Fraud Prediction in E-Commerce Transactions using TabNet with Advanced Data Balancing Techniques

Abstract

E-commerce platforms handle substantial financial transactions, making them attractive targets for fraudulent activity. Those activities may include the use of stolen credit cards or the creation of accounts that exploit online systems. These fraudulent techniques are continuously evolving, which makes prediction increasingly challenging. To address this problem, TabNet, a deep learning classification algorithm, has been applied and compared with two standard machine learning algorithms: Support Vector Machine (SVM) and K-Nearest Neighbors (KNN), in predicting fraudulent transactions in ecommerce. Moreover, since the dataset utilized in this study is extremely imbalanced, three advanced
balancing techniques have been used, i.e., SMOTE, SMOTE-ENN, Borderline -SMOTE. Furthermore, because the accuracy evaluation metric is insufficient for evaluating model performance on an imbalanced dataset, three additional metrics (i.e., recall, specificity, and AUC) have been adopted to evaluate the performance of the classifiers and provide a more comprehensive assessment. Accordingly, TabNet with Borderline-SMOTE approach achieved the best overall performance.