Transfer Learning Based Credit Scoring
Qiancheng Wei, Liu Ying, Kaichao Wu · 2021
Financial institutions often encounter cold-start problems when building credit scoring models. This paper applies transfer learning to credit scoring predictions to help solve this problem. According to the characteristics of the credit scoring task, we make the following two innovations. The first one is to propose a SPY-Transfer model. We transform the SPY algorithm in Positive-Unlabeled(PU) field to enable it to select more valuable samples from the source data and fill them into the target data, thus implement a sample-based migration learning method. The second one is to propose a SPY-TrAdaBoost model, which uses the idea of SPY algorithm and the calculation method of TrAdaBoost algorithm to select more suitable samples from the source data to fill in the target data. In the credit score prediction experiment, the transfer learning method proposed in this paper has better performance than the cold-start modeling. In the same task, compared with the classic transfer learning method TrAdaBoost algorithm, the model proposed in this paper also has a greater performance advantage.