| In recent years,China’s cross-border e-commerce has developed rapidly.From the perspective of national policies,China will vigorously promote the development of crossborder e-commerce.With the development of the Internet,credit card payment as a payment method,more and more overseas users purchase the goods they need through using credit cards.Due to the characteristics of cross-border business such as digitalization,convenience and rapid capital flow,credit card theft has emerged in an endless stream,which has seriously damaged the interests of businesses and e-commerce platforms,and affected the stable development of the credit card industry.In the past,the detection of credit card fraud mainly involves machine learning algorithms.Common classification algorithms have high requirements on data,requiring a small difference between the number of fraud samples and normal samples in the data set.However,in real life,the overall proportion of cross-border e-commerce card theft and fraud transactions is very low.It is very important to improve the prediction effect and enhance the recognition rate of small samples.This paper takes M Company as an example to introduce the rule-based transaction fraud risk management measures currently used by the company,and use the actual transaction data of M company to evaluate the prediction effect of expert rules using two indicators,precision rate and recall rate.It is found that the precision rate of blacklist is only 31.91%,and the recall rate is 71.99%.In view of the actual transaction data of M Company,this paper attempts to use the generative adversarial networks(GAN)to deal with the data imbalance and sets up two groups of comparative experiments at the same time.One group compares CNN,SMOTE-CNN,ADASYN-CNN,Random Under Sampler-CNN,and GAN-CNN;The other group compares the results of common classification algorithms,such as Random Forest,SVM,Logistic regression and GAN-CNN,and takes the recall,precision,AUC value,F1 value and Kappa coefficient of the test set as evaluation indicators,and comprehensively evaluated the effects of each model based on the improved TOPSIS method.In addition,based on the effectiveness improvement of the GAN-CNN model and the importance of the variables measured by the random forest algorithm,this paper puts forward suggestions for improvement of the e-commerce platform.The research findings are as follows:(1)The prediction model based on the generative adversarial networks is better than the prediction model based on the traditional unbalanced data processing method.From the empirical results,we can see that the precision rate of the GAN-CNN method is 1,and the recall rate is 0.999632.(2)The recall and precision of the prediction model based on generative adversarial networks have significantly improved compared with the best rule mentioned in the previous experiment,the blacklist rule,with the precision increased by 0.6809 and the recall increased by 0.2797. |
