Abstract
Default detection, a crucial aspect of individual credit scoring, has attracted considerable attention in research. Previous approaches typically focus on classifying applicants using only explicit attributes, overlooking the importance of latent relations among them. Concurrently, graph-based techniques have emerged as promising tools for credit scoring. However, existing graph-based methods often need to be more accurate in aggregating information from limited neighbors, which can lead to misclassification when the target node has differently labeled neighbors. Motivated by these challenges, we propose a Local and Global Information-aware Graph Neural Network (LG-GNN) approach for default detection. By leveraging the loan applicant relation graph, LG-GNN dynamically learns the representation of the target node from both local and global perspectives. Furthermore, it adaptively fuses the information from these perspectives and employs contrastive learning to enhance feature variations. Extensive experiments demonstrate the superiority of LG-GNN over mainstream methods across several widely used default detection datasets. Specifically, LG-GNN achieves an average relative performance improvement of 47.9% compared to baselines. Moreover, compared to the most competitive default detection methods, LG-GNN exhibits an average performance improvement of 11.9%. Our code is publicly available at https://github.com/BERA-wx/LG-GNN.
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have