Construction And Application Of Credit Risk Control Model Based On Deep Neural Network

With the continuous development of the Internet finance industry,the number of users relying on the Internet platform for personal loans is increasing.Because Internet online platform natural opacity,and domestic for the construction of personal credit information system is not perfect,effective forecast users default probability into credit risk control system,the core of building from the user's credit before more effective information in daily behavior data mining,to extract more complex characteristics become more important.The user's original features are sparse and high-dimensional,and many fields are classified fields.The correlation between each field and other fields is not obvious.The manual extraction of complex combination features is heavy and inefficient,and it is impossible to generalize the combination features that have not appeared in the training.DeepFM model based on deep neural network and factorization mechanism can learn the complex correlation of original feature data,and at the same time,learn the low-order and high-order combination features,and effectively correlate the complex features with the results of user repayment performance.This article first builds a DeepFM model on the public Lending Club user loan data set to verify the gain effect of combined feature learning.Compared with the logistic regression model and XGBoost model that have been successfully used in the credit risk control model,it is found that the DeepFM model has an AUC evaluation index.The above performance has gains of 9% and 2%,respectively,proving that the model's advantages are reflected in the automatic learning of low-order and high-order combined features,which is suitable for data sets with multiple classification fields.Secondly,this article improves the basic DeepFM model.Because traditional factorization machines(FM)lack the ability to distinguish the importance of combined features,in order to learn the combined feature weights more effectively,an attention mechanism is introduced on the basis of the basic model.Then add the XGBoost model as an automatic feature selection tool,filter the features set above the threshold of feature importance,and input the DeepFM model for training.After testing and comparison,it is found that the improved model improves the performance indicators tested on the data set on the basis of the original model.After the features are filtered,the input feature dimension is reduced,and the model training efficiency is also improved.Finally,this paper will implement the improved credit risk control model based on real user loan data,explain the scale of the data set and the reliability and validity of the data set,and verify the application effect of the improved model.The design and implementation of the risk control model online test system,the system can complete the feature selection,model training,model online,model test functions,the feature selection method and the construction of credit risk control model is applied to the actual system implementation.Figure 41,Table 24,Reference 33.