基于LeNet-5模型和门卷积神经网络的信用评分模型实证研究
发布时间：2019-05-12 15:21
【摘要】：信用评分是基于客户信用等级的数值表达式,是用于评估和防范违约风险的有用工具,也是信用风险评估中的一种重要方法。因为风险具有客观存在性和控制艰巨性,为了减小信用风险事件带来的损失,需要选取合适的信用评分模型来进行信用风险控制。建立信用评分的方法有很多,本论文对信用评分模型的研究现状进行了文献综述。对常用的评分方法包括线性回归、判别分析、贝叶斯网络、逻辑回归、模糊逻辑、决策树、支持向量基、遗传算法、神经网络等。其中,神经网络模型有着更强的非线性处理能力,能提高信用评分模型的精度,其次本论文考虑构造一个新的神经网络信用评分模型。卷积神经网络(CNN)是一个多层的神经网络,通过在相邻层的神经元间加入卷积运算,来提取数据的特征。卷积神经网络有诸多类型,其中LeNet-5卷积神经网络是经典的卷积神经网络。LeNet-5具有权值共享和最大池化的特征,其识别效果显著。其网络结构中卷积层与最大池化层交替堆栈是LeNet-5网络的核心。门卷积网络(GCNN)是由Facebook人工智能实验室提出的深度学习模型,在训练方面表现出优异的性能。卷积神经网络通过卷积核权值共享,对输入进行整体建模。递归神经网络(RNN)是带有循环的神经网络,将过去的信息可以保留在系统中。网络具有记忆功能,能保留之前计算过的一些信息,并用在之后的计算中,但标准的递归神经网络无法对长期依赖进行学习。而长期记忆网络(LSTM)是一种特殊的递归神经网络,能够避免长期依赖问题。所有的递归神经网络都有链状的重复神经网络模块。LSTM有四个以特殊方式交互的神经网络层来取代单一的神经网络层。再运用门控机制对信息进行选择。门卷积网络模型通过将引入LSTM的门控机制引入到卷积神经网络中,实现对输入通过由局部到整体的共享权重而进行的全局建模,并用门机制对信息进行识别和判断,取得了良好的效果。本文的主要工作包括:以GCNN在卷积神经网络中加入输入门控制信息的思想为基础,参照CNN的经典模型LeNet-5的结构,将GCNN与LeNet-5模型相结合,并对层结构特征进行改进,发挥两个模型的优势,提高神经网络的优化能力,结合个人信用风险的特点,构造新的信用评分模型。采用知名P2P机构的借款人用户信息作为样本,对基于GCNN和LeNet-5的新模型进行训练和测试,用多分类支持向量基作为对比实验,实验结果表明该模型实证方面具有良好的效果。
[Abstract]:Credit score is a numerical expression based on customer credit rating. It is a useful tool for evaluating and preventing default risk, and it is also an important method in credit risk assessment. Because the risk has the objective existence and the control difficulty, in order to reduce the loss caused by the credit risk event, it is necessary to select the appropriate credit score model to carry on the credit risk control. There are many methods to establish credit scoring. In this paper, the research status of credit scoring model is reviewed. The commonly used scoring methods include linear regression, discriminant analysis, Bayesian network, logical regression, fuzzy logic, decision tree, support vector basis, genetic algorithm, neural network and so on. Among them, the neural network model has stronger nonlinear processing ability and can improve the accuracy of the credit scoring model. Secondly, a new neural network credit scoring model is considered in this paper. Convolution neural network (CNN) is a multi-layer neural network. By adding convolution operation between neurons in adjacent layers, the characteristics of data are extracted. There are many types of convolution neural networks, among which LeNet-5 convolution neural network is a classical convolution neural network. Lenet-5 has the characteristics of weight sharing and maximum pooling, and its recognition effect is remarkable. The alternating stack of convolution layer and maximum pool layer in its network structure is the core of LeNet-5 network. Gate convolution network (GCNN) is a deep learning model proposed by Facebook artificial intelligence laboratory, which shows excellent performance in training. The convolution neural network shares the weight of the convolution kernel to model the input as a whole. Recurrent neural network (RNN) is a neural network with cycle, which can keep the past information in the system. The network has the function of memory and can retain some of the previously calculated information and be used in the subsequent calculation, but the standard recurrent neural network can not learn from the long-term dependence. Long-term memory network (LSTM) is a special recurrent neural network, which can avoid the problem of long-term dependence. All recurrent neural networks have chain repetitive neural network modules. LSTM has four neural network layers that interact in a special way to replace a single neural network layer. Then the gate control mechanism is used to select the information. By introducing the gate control mechanism introduced into the convolution neural network, the gate convolution network model realizes the global modeling of the input through the shared weight from the local to the whole, and uses the gate mechanism to identify and judge the information. Good results have been achieved. The main work of this paper is as follows: based on the idea that GCNN adds input gate control information to convolution neural network, referring to the structure of CNN's classical model LeNet-5, GCNN and LeNet-5 model are combined, and the layer structure characteristics are improved. Give full play to the advantages of the two models, improve the optimization ability of neural network, combined with the characteristics of personal credit risk, a new credit scoring model is constructed. The borrower user information of well-known P2P institutions is used as a sample to train and test the new model based on GCNN and LeNet-5, and the multi-classification support vector base is used as a comparative experiment. The experimental results show that the model has a good empirical effect.
【学位授予单位】：深圳大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：F832.4
本文编号：2475481
[Abstract]:Credit score is a numerical expression based on customer credit rating. It is a useful tool for evaluating and preventing default risk, and it is also an important method in credit risk assessment. Because the risk has the objective existence and the control difficulty, in order to reduce the loss caused by the credit risk event, it is necessary to select the appropriate credit score model to carry on the credit risk control. There are many methods to establish credit scoring. In this paper, the research status of credit scoring model is reviewed. The commonly used scoring methods include linear regression, discriminant analysis, Bayesian network, logical regression, fuzzy logic, decision tree, support vector basis, genetic algorithm, neural network and so on. Among them, the neural network model has stronger nonlinear processing ability and can improve the accuracy of the credit scoring model. Secondly, a new neural network credit scoring model is considered in this paper. Convolution neural network (CNN) is a multi-layer neural network. By adding convolution operation between neurons in adjacent layers, the characteristics of data are extracted. There are many types of convolution neural networks, among which LeNet-5 convolution neural network is a classical convolution neural network. Lenet-5 has the characteristics of weight sharing and maximum pooling, and its recognition effect is remarkable. The alternating stack of convolution layer and maximum pool layer in its network structure is the core of LeNet-5 network. Gate convolution network (GCNN) is a deep learning model proposed by Facebook artificial intelligence laboratory, which shows excellent performance in training. The convolution neural network shares the weight of the convolution kernel to model the input as a whole. Recurrent neural network (RNN) is a neural network with cycle, which can keep the past information in the system. The network has the function of memory and can retain some of the previously calculated information and be used in the subsequent calculation, but the standard recurrent neural network can not learn from the long-term dependence. Long-term memory network (LSTM) is a special recurrent neural network, which can avoid the problem of long-term dependence. All recurrent neural networks have chain repetitive neural network modules. LSTM has four neural network layers that interact in a special way to replace a single neural network layer. Then the gate control mechanism is used to select the information. By introducing the gate control mechanism introduced into the convolution neural network, the gate convolution network model realizes the global modeling of the input through the shared weight from the local to the whole, and uses the gate mechanism to identify and judge the information. Good results have been achieved. The main work of this paper is as follows: based on the idea that GCNN adds input gate control information to convolution neural network, referring to the structure of CNN's classical model LeNet-5, GCNN and LeNet-5 model are combined, and the layer structure characteristics are improved. Give full play to the advantages of the two models, improve the optimization ability of neural network, combined with the characteristics of personal credit risk, a new credit scoring model is constructed. The borrower user information of well-known P2P institutions is used as a sample to train and test the new model based on GCNN and LeNet-5, and the multi-classification support vector base is used as a comparative experiment. The experimental results show that the model has a good empirical effect.
【学位授予单位】：深圳大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：F832.4
【参考文献】
相关期刊论文 前1条
1 庞素琳;概率神经网络信用评价模型及预警研究[J];系统工程理论与实践;2005年05期
，本文编号：2475481
本文链接：https://www.wllwen.com/jingjilunwen/huobiyinxinglunwen/2475481.html
