Recently, a variety of machine learning methods has been proposed and achieved significant improvement in a broad area. In the credit scoring task of P2P(Peer-to-peer) lending, these methods also have been applied by many researchers. However, it is because of the nature of default prediction tasks, models which deal with high dimensions, have not applied yet (ex. self-attention). This paper attempts to apply self-attention & dynamic convolution to credit scoring task in P2P lending. The model’s best empirical results are 99.54% and 0.997 for 2 class and 99.59% of accuracy and 0.9972 of AUC for 3 class classification, the models outperform previous research models in 3 class classification by 6%. During experiments, the models need much small epochs comparing to previous best model. Then, conduct further less sample experiments and confirm improvement. The best result is resulted in 99.92% of accuracy and 0.9994 of AUC scores. The models are robust to imbalanced data.
