Empirical Study Of Credit Scoring Model Based On The LeNet-5 Model And Gated Convolutional Networks

Credit scoring is a numerical expression based on customer credit rating.It is a useful tool for assessing and preventing default risk.It is also an important method in credit risk assessment.Because the risk is objective and it is difficult to control,in order to reduce the loss caused by the credit risk,we need to select the appropriate credit scoring model to control the credit risk.There are many methods to establish the credit scoring model.In this thesis,first the status of the credit scoring models is reviewed.The commonly methods used in scoring include linear regression,discriminant analysis,Bayesian network,logical regression,fuzzy logic,decision tree,support vector method,genetic algorithm,neural network and so on.Among them,the neural network model has a stronger nonlinear processing ability,that can improve the accuracy of credit scoring model.Next,this paper considers to build a new construction of a neural network to credit scoring.Convolutional neural network(CNN)is a multi-layer neural network that extracts the characteristics of data by adding convolution operations between neurons in adjacent layers.There are many types of convolution neural networks,in which the Le Net-5convolution neural network is a classical convolution neural network.Le Net-5 has the characteristics of weight sharing and max pooling,and its recognition effect is remarkable.The stacking of the convolution layers and the max pooling layers in the network structure is the core of the Le Net-5 network.Gated Convolutional Neural Network(GCNN)is a deep learning model proposed by Facebook Artificial Intelligence Lab,which shows excellent performance in training.The convolution neural network constructs the input by convolution.Recurrent Neural Network(RNN)is a neural network with loops,and the past information can be retained in the system.The network has a memory function that retains some of the information that was previouslycalculated and would be used in subsequent calculations,but the standard recurrent neural network can not learn for long-term dependencies.The long short term memory network(LSTM)is a special recurrent neural network that avoids long-term dependency problems.All recurrent neural networks have chained repetitive neural network modules.LSTM has four neural layers that interact in a particular way to replace a single neural layer.And then use the gating mechanism to select information.The gated convolution network model introduces the gating mechanism into the convolution neural network to realize the global modeling of the input through the shared weight from local to global,and uses the gating mechanism to identify and judge the information.The main contributions of this thesis are as follows: Based on the idea of Gated Convolutional Networks adding the gating information mechanism in the convolutional neural network,referring to the structure of the classical convolutional neural network model Le Net-5,this thesis combine the Gated Convolutional Networks with Le Net-5model to improve the advantages and the optimization ability of the two models.The new credit scoring model is constructed according to the characteristics of personal credit risk.A new model based on GCNN and Le Net-5 is trained and tested by using the borrower user information of well-known P2 P institutions.The experimental results show that the propose model has a good performance.