A case study in credit fraud detection with SMOTE and XGboost

Abstract

Credit fraud observations are minority in the sample set, variables tend to be seriously unbalanced, and the prediction results tend to be biased towards more observed classes. Common resolution usually constructs 1:1 data, either cutting off part of more classes (undersampling) or reducing classes for bootstrap sampling (oversampling). XGBoost is an efficient system implementation of Gradient Boosting, and also GB algorithm based on CART. Based on the real online transaction data of an Internet financial institution, this paper studies the performance of XGBoost algorithm on the original data set, the undersampling and SMOTE data sets respectively.