Research On Credit Scoring Model Based On Imbalanced Data Sampling And Convolutional Neural Network

The booming consumer credit market can not only promote the economic development,but also bring new challenges to the financial system.As an effective personal credit risk assessment tool,the credit scoring model can analyze the applicant's default risk by investigating the candidate's credit data and help financial institutions reduce the loss.Usually,the credit scoring datasets in the practical applications are unbalanced.What's more,most of the existing scoring models are shallow structures with weak feature extraction ability.Taking the imbalanced characteristics of personal credit scoring datasets and the shortages of the existing credit scoring models into consideration,the research contents are as follows:(1)In this thesis,an integrated sampling method is designed to solve the problem caused by the imbalanced datasets in personal credit scoring field.Firstly,new minority samples are generated by the KmeansSMOTE algorithm.At the same time,the redundant samples in majority class are removed by the RBU algorithm.Finally,the processed minority and majority class samples are put together to obtain a balanced dataset.The experiment result on several classifiers shows the proposed method can effectively improve the performance of the classifiers.(2)A personal credit scoring model based on 1DCNN(One Dimension Convolutional Neural Network)and ELM(Extreme Learning Machine)is proposed in this thesis.Firstly,in order to solve the problem that the feature extraction ability of shallow architecture credit scoring model is not strong enough,the 1DCNN is used to extract the hidden features of the datasets.Then,the extracted features are transferred to the ELM for model training.Finally,the trained ELM model is applied to classify the data.Four realistic datasets are used for test,the experimental results show that multiple indexes of the 1DCNN-ELM model proposed in this thesis are better than several existing models.