Research And Implementation On Customer Credit Classification Based On SVM

With the vigorous development of the consumer credit market,the size of the user and the amount of data has soared these years.Manual credit risk Assessment for all users cannot meet growing market demands.Therefor technologies of credit risk assessment based on machine learning are increasingly popular,and Support Vector Machine(SVM)model,one of the best,provides an efficient,fast intelligent discriminant method for financial institutions.However,the accuracy and training speed of the model need to be improved.In this paper,the parameter selection process and incremental learning process of SVM are analyzed,studied and improved.(1)Considering the different economic loss caused by rejecting a loan applicant with good credit and agreeing a loan applicant with bad credit,this paper introduces type I error and type II error to propose weighted harmonic mean to take the place of classification accuracy as new model evaluation indicator.Besides,a risk coefficient variable is designed to make a compromise about two types of error.The results show that the weighted harmonic mean can better reflect the expected profit and loss of credit institutions,meanwhile,and meet the individual needs of different credit institutions.(2)An SVM parameter selection model based on hybrid bat algorithm is established to explore choice of parameters of SVM model,for there is no fixed method and theory.penalty factor C1,C2 and the kernel parameters γare reflected by the position of individual bats,and the fitness values of individuals are represented by weighted harmonic mean.By introducing the differential evolution process,crossover,selection and mutation behavior occur when bat population are during the iterative process,which enhances the global search ability of the algorithm.The experimental results show that the SVM parameter selection model can improve the robustness and convergence accuracy,and effectively alleviate the premature convergence problem.(3)An improved SVM algorithm based on the weighted candidate set is proposed to cope with long training time and bad results.With regard to the history training set,the selection of training sample points is added to the candidate set based on the combination condition of the hypersphere and the classification surface.As for the incremental sample set,the improved KKT(Karush-Kuhn-Tucker)condition is used to screen incremental samples,selecting all violation and part samples near the Optimal classification hyperplane.Based on the analysis and experiments,the SVM incremental learning algorithm based on candidate set can effectively reduce the number of samples involved in training and preserve the potential support vector set at the same time.