A Tabnet based Card Fraud detetion Algorithm with Feature Engineering
Lei Zhang, KaiFeng Ma, Fang Yuan, WenJun Fang · 2022 2nd International Conference on Consumer Electronics and Computer Engineering (ICCECE) · 2022
With the development of technology, there are some inventions has been raised, such as credit cards for paying in electronic transaction. People can trade in a short time with credit cards. However, it will be a double-edged sword because it provided a chance for criminals, and in turn, increased fraud rate. Therefore, there is an urgent need for researchers to research and proposed some methods. In our article, we utilize TabNet as our model. The dataset is provided by IEEE-CIS and it contains many transaction records and whether the transaction is fraud. In the feature engineering process, we compress the datasets and fill the missing value with -999. In the training process, the cross validation has been used, we split the dataset into 5 folds, and train them one by one. And we do some improvement to compress the memory. Finally, to evaluate how well TabNet performed, we compare it with other classical models: Naïve Bayes and XGBOOST. The result shows that our model has the best performance using the metrics ----accuracy and AUC-ROC score among these models. It is obvious that TabNet owns the highest AUC-ROC score and accuracy, and they are 0.884 and 0.965 respectively. In the future, we can improve the model and data based on this information to improve the classification effect and apply the model to other fraud detection problems.