Research On The Application Of Migration Learning In Cash Loan Default Prediction

In recent years,with the development of Internet Finance,a new kind of financial products called “payday loan” have developed rapidly.The default forecast of paydayloan users is becoming more and more important,but due to the emerging of paydayloan business and the sensitivity of financial data,it is impossible to obtain enough tagged users for start-up cash lending companies or companies that expand cash lending business.In order to solve the "cold startup" problem of the risk control platform and expand the ideas of traditional default prediction modeling,this paper use "transfer learning" method and a company's credit loan data and payday-loan data to establish a default prediction model,study the feasibility of transfer learning in the field of risk control.Firstly,we reviewed the current research and application in the field of default control,and introduced related theory about the algorithms using in this paper.Then we introduced the characteristics of the data,preprocessed the data and made feature engineering.The traditional model is established by the algorithms of Logistic Regression and XGBoost,and the effect of two models are analyzed and evaluated.Secondly,we used the credit loan data as the source data and the cash credit data as the target data,establishing prediction model through TrAdaboost algorithm and the transfer learning framework based on XGBoost sample filter and K-means,then compared transfer method with traditional method.The results showed that the transfer learning algorithm can be useful,the prediction ability of model is improved.At the same time,the source domain sample screening method based on XGBoost classifier proposed in this paper helps to improve the effect of the payday-loan default prediction model.At the end,based on the model effect and specific business characteristics,this paper proposes the methods and methods of the transfer learning model applied in the payday-loan business.Through in-depth analysis of model effects and our specific business,this paper believes that the transfer learning method is helpful for the ‘cold startup' situation without large amount of user sample data.The transfer learning methods have certain application ability in the construction of payday-loan risk control platform.