Improved credit card fraud detection method based on XGBoost algorithm

Abstract

With the development of the Internet and technology, credit cards are more widely used and transaction data are larger. The data set of credit card fraud is a typical imbalanced data problem. The model should ensure that fraud is detected and customer service quality is guaranteed. Improving both the precision and the recall rate is the focus of current research. However, when the precision is constrained by the current level of machine learning, it is a good choice to use different models to evaluate and obtain more data for manual use. Few papers use the results of both models for comprehensive judgment. In this paper, two sampling methods (undersampling-NearMiss, oversampling-SMOTE) and three algorithms (Logistic, Neural Network, XGBoost) are used to analyze data based on the transaction records in Europe within two days in September 2013. The above six cases were compared to explore appropriate detection methods. The study found that Nearmiss-XGBoost had the best recall rate, SMOTE -XGBoost had the best comprehensive result and precision. Compounding the results of the two models can improve the precision and recall.