Research On Credit Risk Prediction Based On Improved DenseNet-BC

In the context of financial marketization,both emerging Internet finance and traditional offline financial institutions regard credit business as their main business and source of profit.Although financial institutions can get considerable profits from it,they will bear various risks caused by the borrower's default if they cannot make accurate prediction of the borrower's credit risk.With the constant change of the global economic environment and the rapid growth of customer data,the traditional credit risk prediction model is no longer applicable to the current objective situation.How to improve the accuracy of credit risk prediction and build a reliable risk prediction model has become an important issue faced by the industry and academia.This paper studies the problem of credit risk prediction and puts forward a credit risk prediction model based on improved DenseNet-BC.The main work of this paper is as follows:(1)Modern economic data sets usually have high and unbalanced problems,which leads to the decrease of training accuracy and poor robustness of credit risk prediction models.Aiming at the above problems,a DenseNet-BC credit risk prediction model incorporating Focal Loss is proposed.Based on DenseNet-BC as the basic model architecture,this model introduces Focal Loss Loss Function in the classification layer and utilizes the dense connection characteristics of DenseNet-BC.It not only helps to deepen the number of network layers to enhance the effective extraction of borrower characteristic information,but also can effectively combine the important characteristic information of shallow network and deep network.At the same time,Focal Loss function adjusts the category imbalance between samples and the quantity imbalance between difficult and easy samples through weight,reduces the error caused by multi-proportion category samples and easily classified samples on model prediction,and improves the generalization of the model.(2)For the DenseNet-BC algorithm model integrating Focal Loss,the influence of learning rate on training results in the training process of deep neural network model due to the high-dimensional sparse matrix formed by dumb quantization of discrete features.Furthermore,a DenseNet-BC credit risk prediction model combining the Auto Encoder and the static restart SGD is proposed.The high-dimensional discrete feature vectors processed by dummy variables are extracted by an Auto Encoder,and then the extracted discrete feature vectors are combined with the preprocessed continuous feature vectors as the input of the improved DenseNet-BC algorithm model.At the same time,the static restart SGD is used in the model training process.The algorithm can effectively guarantee under the condition of the reduce amount of calculation and memory loss better characteristics,static restart SGD can make the model in the process of training to the adaptive learning rate update at the same time,so that the learning rate is not easily affected by the size of the initial learning rate,and avoid the model falling into local optimum due to too small learning rate during the training process or missing the global optimal due to too large learning rate,so as to improve the accuracy of the algorithm model prediction.The algorithm in this paper is based on the credit risk prediction data set published by Lending Club.It is compared with a variety of traditional algorithms and DenseNet-BC,and the experimental results show that the algorithm in this paper has certain advantages.