Research On Credit Decision-Making Based On Deep Reinforcement Learning

The risk analysis of credit customers in financial institutions is a typical classification problem,and the problem of data imbalance widely exists in the decisionmaking of credit customers’ default risk,which affects the effective identification of risky customers.At present,domestic and foreign commercial banks have widely used machine learning and deep learning in credit risk analysis,but both classification models have shortcomings in the classification scenario of unbalanced data facing the credit field.Although the conventional machine learning model has simple structure and strong interpretability,it is widely used in the financial industry,but the model effect is not as good as the deep learning classification model.Although deep learning has strong classification performance,its internal classification process is invisible,it can’t explain the control of the algorithm operation process to customers,it lacks interpretability,and it is easy to over-fit.In view of the problems existing in the application of machine learning and deep learning in the field of credit risk analysis,this paper proposes an improved Easyensemble combination model,which is based on the preprocessing of boundary oversampling and can effectively improve the prediction accuracy,thus better controlling the credit default risk.In the combined model,the sample formation ratio of Borderline-Smote is controlled,and the K-value optimization algorithm is added to improve the accuracy of data processing.The improved preprocessing algorithm is applied to the generation of new minority samples,which are input into the Easyensemble classifier to obtain the final classification result.In order to further improve the interpretability of imbalanced data set classification in the scenario of credit default prediction,this paper considers the design of imbalanced data classification model by integrating deep reinforcement learning,modeling the integration problem of multiple weak classifiers as a sequential decision-making problem,and then introducing reinforcement learning framework to build a data-totask credit risk prediction model,and studying the ability of deep reinforcement learning to solve credit risk prediction problems,thus improving the early warning ability and credibility of the model.Experiments on four kinds of customer default data sets in the open credit field show that compared with the traditional imbalanced data classification algorithms,such as Easyensemble,XGBoost,Bagging and Blance Cascade,the F1-score value and AUC value of DDKBSE in dealing with extremely imbalanced data sets Credicard and Bank are the highest among the five groups of algorithms.The above results show that in the bank’s credit default risk prediction,using the Easyensemble combination model based on boundary oversampling pretreatment and deep reinforcement learning can improve the model performance of dealing with extremely unbalanced credit data.The purpose of this paper is to explore the credit default risk prediction method based on Easyensemble combination model and deep reinforcement learning framework,in order to provide more accurate personal credit default risk prediction for commercial banks,thus improving the early warning level of commercial banks’ risk control system and promoting the development of domestic credit industry.