Research On Credit Scoring Model Based On Oversampling FBSMOTE

With the rapid development of Internet finance,the demand for personal credit has also increased.While enjoying the opportunities of rapid development of credit business,financial institutions are also faced with the challenges of more complex and diversified credit scoring.Aiming at the class imbalance problem of credit data and the complex distribution of default data,this thesis proposes a class balance processing method and a credit scoring model.Firstly,for the class imbalance problem,this thesis proposes a new oversampling method FBSMOTE.When considering the distribution of knearest neighbors of minority class samples,the method makes full use of the class purity information of k-nearest neighbors and the distance information between samples to calculate the weighted similarity and dissimilarity ratio of minority class samples.It becomes a minority class sampling probability that pays more attention to the classification boundary,and finally generates a new minority class sample by random linearity.This thesis conducts experiments on public data in multiple fields,and combines multiple classification models with other oversampling methods to conduct comprehensive analysis on multiple evaluation indicators.It is verified that FBSMOTE effectively improves the accuracy of the classification model.Secondly,based on the idea of FBSMOTE and ensemble model,this thesis uses FBSMOTE to improve the classification accuracy of the base classifier,and proposes an ensemble model FBSMOTEBoost to deal with the class imbalance problem.This thesis verifies the effectiveness and feasibility of FBSMOTEBoost through comprehensive analysis by conducting experiments on 13 public datasets.Finally,on the problem of credit default prediction,the credit scoring model FBSMOTE-ODM is proposed;FBSMOTE is used to solve the problem of class imbalance,and the optimal interval distribution machine based on the construction of the optimal hyperplane of class distribution is used to improve the classification accuracy and enhance the model’s performance and robustness.Through comprehensive analysis on the real credit data set,it is verified that FBSMOTE-ODM has a better ability to identify default data.