Research On Payday Loan Overdue Forecast Based On Transfer Learning

Relying on the innovation of Internet technology and the transformation of payment methods,the number of Payday loan products has ushered in explosive growth,and the problem of overdue forecast of Payday loan users has also become a hot research topic.Traditional risk control algorithms require the support of sufficient data to get the ideal model effect.However,due to the late start of Payday loan business in China,the data accumulation is insufficient.Moreover,credit data is sensitive and cannot be obtained from public channels,which leads to the cold start problem of Payday loan business.Therefore,models built based on traditional machine learning methods tend to be under-fitting,with poor model accuracy and stability.Transfer learning is a good idea to solve the cold start problem.The idea is to use the knowledge learned on another data-rich related domain to enhance the learning effect of the target domain.Therefore,this paper introduces transfer learning into the study of Payday loan overdue forecast problem,allowing traditional credit data with sufficient samples to assist the construction of Payday loan models,and exploring the feasibility and effectiveness of transfer learning in the Payday loan business.In this paper,two transfer learning models are innovatively proposed from two perspectives.Firstly,from the perspective of instance based transfer learning,through in-depth study of the similarity between the PU learning problem in semi-supervised learning and the instance based transfer learning,and drawing on the idea of two-stage algorithm in PU learning and Spy algorithm,a two-stage transfer learning method based on Spy algorithm is proposed.The first stage of the algorithm is the screening of similar samples.The Spy algorithm is used to find the most similar samples in the traditional credit data to the Payday loan data,so as to expand the training set.The second stage of the algorithm is to build a classifier on the new expanded training set.Secondly,from the perspective of model based transfer learning,this paper combines the idea of ensemble learning with model based transfer learning,and puts forward a transfer learning model based on Stacking.The method trains models on Payday loan data and traditional credit data independently,and then treats the transfer strategy itself as the learning object,using the Stacking algorithm to combine the prediction results of the source domain model and the target domain model.To test the validity of the proposed two transfer learning models,this paper takes the traditional credit data and Payday loan data of a fintech company as empirical objects,and sets up two sets of benchmark models for comparative analysis.The empirical result shows that after introducing different transfer learning methods,the performance of the model has been improved to varying degrees.The research in this paper has positive significance for Payday loan business,by introducing semi-supervised learning and ensemble learning into the field of transfer learning,two transfer learning models are innovatively proposed to enrich the transfer learning methods and provide a new solution to the cold start problem of Payday loan business.