Research On Credit Default Identification Method Based On Deep Learning

Risk control is at the heart of the financial field.With the development of Internet finance and the arrival of the era of big data,the number of loan transactions and the amount of loans have also increased significantly,which makes financial risk management very important.However,traditional machine learning models encounter bottlenecks in training and forecasting when dealing with massive,high-dimensional,and complex data.In recent years,deep learning technology has solved the shallow defects of neural networks and can construct multi-layer nonlinear relationships.Therefore,it has been widely used in many fields.It is of practical significance to reduce the risk of credit default through deep learning based methods,strengthen financial supervision,and reduce the losses of financial institutions.This paper starts with two aspects of supervised and semi-supervised learning,and constructs two credit default recognition models based on deep learning.The research difficulty in identifying credit defaults is the lack of annotation data for real-time massive transaction data.In view of the fact that the Deep Belief Network(DBN)of the semi-supervised deep learning model requires less tag data and can fully utilize a large amount of unlabeled data to achieve effective feature extraction,this paper proposes a model based on semi-supervised deep learning.Considering the limitations of the traditional unsupervised learning Isolation Forest(iForest)in processing high-dimensional data and its advantages in anomaly detection,a semi-supervised credit default recognition model based on DBN-iForest is proposed.After that,through the simulated annealing algorithm and particle swarm optimization algorithm,the optimization of the main parameters of the isolated forest algorithm is realized,which further improves the classification performance and accuracy of the model.Secondly,because the traditional machine learning-based credit default recognition model has low accuracy in classifying high-dimensional unbalanced financial transactions,this paper proposes a kind of supervised credit default identification based on Convolutional Neural Network(CNN)and Random Forest(RF)algorithm.The model enables the identification of fraudulent transactions from high-dimensional unbalanced credit financial transaction data.The core idea of the algorithm is to construct a two-stage working model CNN-RF.Firstly,the feature extraction of transaction data is automatically performed by constructing a convolutional neural network,and then the classification prediction is performed by a random forest algorithm.The model makes full use of the advantages of CNN's feature extraction for complex data and the generalization ability of random forest algorithm,and achieves good results.Finally,this paper uses the Loan-Default-Prediction provided by Kaggle game as the data set to verify the two models proposed in this paper.The experimental results show that the proposed model can obtain better classification performance.