Research On Credit Card Fraud Detection With Deep Forest

Big data,artificial intelligence and other technologies have promoted the great development of financial payment.The increasing use rate of credit card has brought benefits and convenience to society.Undoubtedly,Credit card payment is rapidly popularized in China while the cases of credit card fraud are also on the rise at the same time.China's annual losses alone exceed 10 billion dollars,which seriously hinders the long-term development of the financial industry.In the face of the increasingly serious credit card fraud crime,the credit card fraud detection system is confronting huge detection pressure.How to prevent the fraudulent transactions as much as possible has become the key work of industry and regulatory departments.With the development of data mining,machine learning and other technologies,which has been used to solve the problems of fraudulent transactions and a number of research outcomes have been achieved.Due to the low frequency and great harm of credit card fraud transactions.This paper expounds the fraud detection problems of credit card and the fraud detection models and algorithms at home and abroad.The main difficulties of the research are as follows: 1.The fraud transaction of credit card only accounts for a very small part of all transaction data,the distribution of data categories presents a highly unbalanced feature,and the positive and negative samples are highly overlapped in the feature information,which leads to the difficulty of distinguishing and wrong classification.The classification algorithm can not effectively identify fraudulent transactions;2.The data generally has the problem of concept drift and lack of adaptability.Due to the change of cardholder's consumption concept and habits,in addition,illegal elements often change fraud strategies to deal with supervision,which makes the fraud detection system unable to adapt to new characteristics and rules.Based on the above analysis,the research contributions are as follows:(1)Focusing on the existing problems of credit card fraud detection and previous research results,with the combination of summarizing the advantages and disadvantages of oversampling and undersampling,a new comprehensive sampling algorithm is proposed in this paper.The Gaussian mixture model and k-nearest-neighbor are used to undersampling the majority of data,then the ADASYN oversampling technique is used for the rest of minority samples,and finally merging two parts of data to get a new data set.The experimental results show that the comprehensive sampling algorithm proposed in this paper can improve the AUC and g-mean,which indicates that the algorithm has good performance.(2)Deep forest algorithm has advantages in representation learning and context understanding,based on the deep forest,we added some basic learners,and we used AUC as the standard to evaluate the cascade performance and control the cascade growth,and proposed SCForest,which is a scalable cascade structure.Combining SCForest with the new sampling algorithm,a credit card fraud detection model is constructed.The experimental results demonstrate that the model has excellent comprehensive performance compared with other models in credit card fraud data set.