Research On Combined Model In Personal Credit Evaluation

After entering the 21 st century,personal credit loan has become a very popular way of national consumption.In addition to the traditional commercial banks,some qualified Internet financial institutions have also entered the field,which makes the personal credit market in China develops to the diversified direction.At present,the methods of personal credit evaluation mainly involve statistics,operational research,non-parametric analysis and artificial intelligence.The single model based on one method has been widely used in the industry.However,due to the limited promotion space of single model,combined model has become a new research hotspot.A reasonable combination of the single classifiers can overcome the shortcomings of the single model,and also make the base classifiers complementary to each other.Therefore,single model is to compare with the combined model in this paper.The research is based on logistic regression and SVM algorithm.First step is preprocessing dataset,such as using KNN to fill the missing value,identifying and dealing with the abnormal value,standardizing the dataset.Secondly,SMOTE algorithm is used to deal with the class imbalance problem.Next,two single models and two combined models are built.Two combined models are ensembled in serial ways and parallel ways respectively.In the process of constructing the combined models,SVM uses the method of sigmoid transformation to obtain the posterior probability,and then combines it with logistic regression.Through empirical analysis,the result shows that combined models have more advantages than single models,and the parallel combined model works best.Compared with single models,combined models not only improve overall accuracy,but also control the error which classifies defaulter as non-defaulter.The improvement of such errors is of great value in credit evaluation.For single model,SVM works better.For the term of robustness,these models all have good stability,for little difference between training samples and testing samples.Among them,SVM's robustness is slightly lower than other models,while logistic regression is the greatest.