Research On Credit Card Fraud Identification Based On The Feature Combination Of GBDT

With the development of economy,people's consumption level has been improved,the demand for convenient mobile payments such as WeChat and Alipay is also more and more vigorous,and people's consumption concept is also gradually transformed.Therefore,credit cards have gradually become an indispensable credit products in people's work,study,and entertainment but with the problem of credit card fraud.Thus,the identification of credit card users belonging to fraudulent transactions has become a top priority.This paper presented a descriptive analysis of 280,000 transactions on European cardholder within two days.Based on two kinds of samples such as normal transaction and fraudulent transaction,the density histogram and cumulative experience distribution diagram of each feature were drawn to study the distribution of feature and K-S test was used to quantify this difference.While K-S test,SVM-RFE and random forest were used for feature selection.The final feature were selected by voting based on the feature selection results of the above three methods.And the Logistic regression model,which was evaluated with Precision,Recall and the area under the receiver operation characteristic curve,was established to identify the credit card fraud.And Recall was increased by 33.4%,F1 by 24.5%,and AUC by 2.2% after the use of SMOTE.Then an improved Logistic regression model was constructed by using GBDT,which is called GBDT+LR.The use of GBDT+LR leaded to more identification of fraudulent card transactions,with Recall,F1 and AUC increased by another 6.1%,3.6% and 0.8%,respectively.Finally,according to the relationship between Precision and Recall on the opposite growth trend under different thresholds and combined with AUC indictor,suggestions for adjusting the anti-credit card fraud intensity were given.