基于类难度感知CDPL的深度学习非均衡信用评价模型OA

Effective management of credit risk is crucial for the stable operation of the financial system.In response to the problem of low accuracy in identifying high-risk default customers in traditional credit evaluation models,a dynamic weight term that reflects the real-time identification difficulty of each sample type is constructed to improve the cross-entropy loss function.On this basis,a back propagation neural network-class difficulty perception loss(BPNN-CDPL)credit evaluation model is established.The default prediction performance of the model is validated using 11 comparative models and five model evaluation metrics.The results show that the proposed model has good adaptability to credit data with different imbalance ratios and credit characteristics,and can significantly improve the identification accuracy of high-risk default customers,which is of primary concern to financial institutions.Compared with existing credit evaluation models,the proposed model exhibits superior default prediction performance.The innovations and contributions of this paper are as follows:introducing dynamic weight terms into the cross-entropy function breaks through the limitations of traditional deep learning models with weak applicability to imbalanced samples,improving the poor recognition performance of the back propagation neural network-focal loss(BPNN-Focal Loss)model on default samples due to not considering differences in sample size between classes,and avoiding the excessive suppression of non-default samples in back propagation neural network-class balanced cross entropy(BPNN-CBCE)model caused by"uniform"sample weighting.This paper extends the theoretical development of deep learning methods and provides new ideas for credit risk management in financial institutions.