Research On Credit Scoring Model Based On Machine Learning

In today's society,the artificial intelligence and Internet industries are developing rapidly,and people's economic consumption mode has gradually changed from traditional consumption mode to credit economy mode.The credit economy mode mainly manifests in personal consumption credit,among which credit card consumption,housing mortgage,automobile loan,education loan and so on belong to personal consumption credit.Under the situation that consumer credit is heating up,a delineation model for personal financial authority,namely the credit scoring model,has been established to protect the security of banks and other financial sectors.The model refers to the credit history of the customer,using machine learning or data mining methods to obtain different levels of credit scores,according to the customer's credit score,to determine the credit line that the customer can hold,thereby ensuring the security of repayment and other services.The credit scoring model mainly evaluates customers' credit.The credit evaluation technology has the characteristics of fast,efficient and standard in handing credit business.Credit scoring model research and development on the one hand can guarantee stable financial order,reduce the cost of credit assessment,improve the consistency of the credit decision,the speed and accuracy,on the other hand can also help improve the efficiency of commercial bank credit work,and expanding consumer credit.Therefore,the development of an effective credit scoring model plays a very important role in stabilizing the development of banking business and reducing the risk of customer credit default.In the existing credit scoring model,the research of domestic and foreign scholars study point focusing mostly ion the traditional machine learning method to improve or to study the characteristic of customer credit,namely feature selection methods.However,based on the traditional machine learning method to build credit scoringmodel for large,irregular credit data,the lack of stability and accuracy.Aiming at the problems existing in the existing credit scoring model,this paper combines grid search,data preprocessing,credit matrix and weighted sampling,feature semantics,mixed group feature selection and other methods to construct a credit scoring model,and predictive analysis of consume credit.The research work of this paper is as follows:(1)Aiming at the problem that the credit scoring model based on traditional random forest algorithm is sensitive to redundant features and the input parameters are difficult to determine,an improved credit scoring model based on random forest is proposed(ICSM-RF).Although the traditional random forest algorithm has the characteristics of simple structure and easy implementation,it also has the disadvantages of uncertain input parameters and sensitivity to redundant noise data in credit data.In view of the above problems,this paper combines the grid search method to optimize the parameters of random forest algorithm and propose the construction of credit matrix and use weighted sampling method to extract the training data.At the same time,in order to improve its prediction accuracy,this paper also adopts corresponding feature selection method to improve the prediction accuracy of credit scoring model based on random forest algorithm.(2)Aiming at the problem that the the credit scoring model constructed by the existing feature selection method has low final prediction accuracy and weak stability between feature subsets,a hybrid credit scoring model based on hybrid clustering and support vector machine is proposed(HC-SVM).Based on the preprocessing of credit data,this paper studies from a multi-dimensional perspective,proposes a new hybrid group feature construction strategy and feature transposition based on feature groups to improve the prediction accuracy of the model and the stability between stable feature subsets.At the same time,the feature semantics is proposed,and the radar map is constructed to explain the credit features to mine more useful information.This paper uses the UCI database and the real credit data set published online to verify and compare it with the existing credit scoring model.The experimental results show that the proposed credit scoring model has better performance in predicting accuracy and stability between subsets of stable features.