Research And Application Of Improved Deep Learning Model Based On DBN

Risk control in the financial field is always the core foundation of the business.We should establish credit scoring model by applying the knowledge of mathematical theory to the understanding of business,solve the credit scoring problem scientifically,reduce the risk of credit default,improve the profit of the enterprise and maximize the profit.The traditional machine learning model brings problems to the training and prediction of the algorithm in the case of large amount of data,complex structure and increasing dimension.The arrival of the era of big data brings problems and challenges to the traditional problem of credit scoring.applying the deep learning algorithm to the credit scoring is a problem worth studying.The main work of this paper is as follows:(1)An improved deep learning algorithm combined with DBN and ELMThe deep belief network(DBN)is a deep learning model with excellent performance,but training needs two stages:pre-training and fine-tuning,which takes a lot of time.The extreme learning machine(ELM)has the characteristics of fast training speed and good generalization performance,but the research shows that ELM often needs more hidden layer neurons to achieve good results,that is,the model needs more memory.In order to effectively combine the advantages of two algorithms,this paper presents a new DBN-ELM algorithm,which combined ELM and DBN.At the same time,in order to effectively utilize the middle hidden layer information,we proposed the algorithm DBN-DELM-ensemble based on bagging.DBN,ELM,IDBN and DBN-DELM,DBN-DELM-ensemble are applied to MNIST and Skin Segmentation datasets of UCI.The results show that the improved algorithm can improve learning speed in ensure the accuracy of existing learning.(2)Use the improved algorithm to model credit score.The modeling of credit scoring requires the analysis of the problems,According to the understanding of the actual business to the mathematical problem,and then combined with the understanding of the data,the data processing and feature mining before modeling,the credit scoring model is established to replace the artificial rules to improve the efficiency.Then the model is evaluated and the performance is bad,and the final model is selected.On the Default of Credit Card Clients dataset and the Give Me Some Credit dataset,modeling predictions are made using traditional method Logistic Regression(LR),Classified Regression Tree(CART),Random Forest(RF),Gradient Boosting Decison Tree(GBDT),it is found that fused model RF and GBDT effects are superior to single model LR and CART.At the same time,the improved model DBN-DELM-ensemble effect is better than the single model LR and CART,but the effect is slightly worse than the RF and GBDT.The result of the model can be used to improve the credit scoring problem.