A transfer learning approach for credit scoring

Credit scoring is one of the major risks faced by banks as well as a significant part of credit risk management. To financial institutions, in most cases, samples of defaults are the minority among total loan samples, while credit client samples making repayments on time are the majority. This phenomenon is called class distribution imbalance problem that is prevailing in credit risk identification. However, existing credit scoring approaches cannot effectively solve the class distribution imbalance problem brought by the scarcity of minority samples. Thus, in this paper, a transfer learning approach is introduced, and the class distribution imbalance problem brought by the scarcity of minority samples is solved through the import of external credit information data. In this paper, a novel transfer learning model is put forward and the classification of target data is facilitated through auxiliary training data transfer so that the efficiency of external credit information using minority samples can be improved. With a new sample initial weight allocation and adjustment strategy, the ability to identify negative samples is highlighted. Through dynamic adjustments to auxiliary training sets, redundant data is duly eliminated as per the pre-set lower weight threshold, reducing the influence of the redundant data on the performance of the classifiers and enhancing the ability of transfer learning to learn imbalanced samples.