Credit Payment Fraud detection model based on TabNet and Xgboot

• Published in: 2022 2nd International Conference on Consumer Electronics and Computer Engineering (ICCECE) pp. 823 - 826
• Main Authors: Cai, Qizhi, He, Jiaxin
• Format: Conference Proceeding
• Language: English
• Published: IEEE 14.01.2022
• Subjects: AUC-Roc score; Computational modeling; Consumer electronics; Credit cards; Data models; Finance; Fraud Detection; memory compressed; Support vector machines; TabNet; Training; XGBoost
• DOI: 10.1109/ICCECE54139.2022.9712842

Nowadays, people tend to use credit cards for their payment efficiency, but credit cards also provide a new opportunity for fraud. Enterprises and researchers have been trying to come up with a method to indicate whether a transaction is fraudulent. In this article, a hybrid model based on the combination of TabNet and XGBoost is proposed. A dataset provided by IEEE-CIS, which contains many transaction records and whether they are fraudulent, is used in this research. The dataset is compressed in the feature engineering process, and the missing values are replaced with −999. In the training process, the cross-validation is used to split the dataset into five-folds then trained one by one. Besides, some improvements are made to cut down the memory usage. Finally, a comparison between TabNet model and two classical models, Naïve Bayes and XGBOOST are made, to evaluate the performance of TabNet. The result shows that our hybrid model has the best performance in terms of two metrics---accuracy and AUC-ROC score---among all models tested. TabNet has the highest AUC-ROC score and accuracy, 0.892 and 0.972, respectively.