Credit Assessment Methods Based On Denoising Autoencoder

In today's society,with the birth of various forms of credit economy,banks and credit agencies have collected a large amount of personal credit data.To make use of these historical data to build credit evaluation models,more and more credit evaluation methods Was presented.At present,credit evaluation methods include Logistic,SVM,decision tree,etc.However,due to the large number of client's attributes involved in the credit evaluation method,many researchers use the collected credit data as the training data of the model.Does not guarantee that the trained model has good generalization ability.Therefore,this paper studies a Denoise Autoencoder neural network that can automatically reduce the dimension of features,and uses it as the basis of credit evaluation model to learn the main features.Then,the learned features are taken as the input of the model Input to improve the model's ability to categorize customers for good or bad credit.The main research work and innovation of this paper include the following contents:First of all,some basic knowledge of credit evaluation is given,the research status of credit evaluation is briefly analyzed,it provides the basis for the new model of credit evaluation.Then,the credit evaluation model based on the combination of Denoise Autoencoder neural network and SVM is described.Here mainly includes the principle of Denoise Autoencoder neural network and its training process,the principle of the credit evaluation method.And for the large amount of data in the credit evaluation,this paper proposes a K-Means clustering-based SVM credit evaluation method.Firstly,the training data is clustered,and then learns the features of each class through the Denoise Autoencoder network,and then uses these learning features to train the SVM model respectively.For unlabeled data,we first calculate the distance from each cluster center and select the model trained by the nearest class to classify them.Finally,the experimental design and verification were carried out.Due to the fact that personal credit data contains a large amount of inaccessible private information,only two public datasets from German credit data and the Kaggle contest were used for experimentation.In this paper,the correctness rate,the misclassification rates of the first and second categories,and the AUC value are used as evaluation standard,and compared with the experimental results of other methods,the correctness and feasibility of the proposed model are verified.