Application Research Of Improved AdaBoost Algorithm In Credit Imbalance Classification Problem

In recent years,with the rapid development of the consumer credit industry,the credit risk challenges faced by banks and other related institutions have gradually increased.In order to control the credit risk,an effective method is needed to correctly identify credit default users,that is,to solve the credit classification problem.However,most credit data sets are imbalanced credit data sets,so the key to controlling the credit risk is to solve the credit imbalance classification problem.In this paper,the AdaBoost algorithm is used to study the credit imbalance classification problem.Firstly,the research results of predecessors are summarized.Then,the related concepts of credit imbalance and the related theories of AdaBoost algorithm are elaborated.Then,based on the sampling method and cost-sensitive idea,a new improved AdaBoost algorithm is proposed.Finally,several different AdaBoost algorithms are used to carry out empirical research on credit imbalance classification problem.This paper uses the personal credit data of a US P2 P network loan company to conduct research.Compared to the traditional AdaBoost algorithm,it is found that the SMOTEBoost algorithm and the RUSBoost algorithm are applicable to the credit imbalance classification problem to a certain degree,but neither of them can effectively improve the accuracy of credit default users.The improved AdaBoost algorithm based on sampling and cost-sensitive proposed in this paper can change the classification performance of the model by adjustingthe small class cost factor,and the best classification performance is obtained when the samll class cost factor is about 6.However,in the improved algorithm,all of the small class cost factor and the number of iterations can not effectively affect the sampling process which is based on the sample weights.In addition,the improved algorithm proposed in this paper is more suitable for the credit imbalance classification problem than the above three algorithms,as long as the value of the small class cost factor is slightly increased,the model with better classification effect can be obtained.