Machine learning models have been actively utilized to quantitatively predict the default probability based on the personal information obtained from loan applicants. Although the relationship between loan applicants is receiving attention as important soft information, only the simple individual network features and their relation have been reflected in the prediction model. In this study, we propose a graph convolutional network (GCN)-based credit default prediction model, which can reflect nonlinear relationships between borrower's attributes and default risk as well as high-order relationships between the borrowers. Three types of information pertaining to the borrowers are separately employed for their relations, namely loan information, credit history information, and soft information. We compare our GCN model to baseline models using the data from an online peer-to-peer lending platform. The results show that our approach outperforms existing classification models and identifies the relative contribution of input attributes.
Bibliographical noteFunding Information:
Funding: This work was supported by the National Research Foundation of Korea (NRF) grant funded by the Korea government (MSIT) [grant number 2020R1A2C2005026].
All Science Journal Classification (ASJC) codes
- Computer Science Applications
- Artificial Intelligence