Enhancing Customer Credit Rating Prediction: A Dual Approach Using Graph Neural Networks

Accurate prediction of customer credit ratings is essential for effective financial risk management. Despite this, recent research in this area has not effectively modeled the relationships between borrowers and guarantors, limiting prediction accuracy. This study introduces a dual-phase model that combines unsupervised and supervised learning to enhance credit rating prediction. In the first phase, unsupervised learning is applied to loan data from customers and guarantors, forming a directed graph where nodes represent individual and corporate customers, enriched with attributes such as customer type and loan history, while edges depict loan guarantee connections, weighted by debt and guarantee frequency. Community detection algorithms identify clusters within this graph, and each cluster is assigned an anomaly score and evaluated against a true credit score threshold to generate initial customer ratings and labels. In the second phase, supervised learning with graph neural networks (GNNs) is used. A Siamese neural network further enhances prediction accuracy, achieving an accuracy rate of 92%. This approach not only improves predictive precision, but also uncovers hidden relationships among customers, providing deeper insights into credit risk and supporting a more robust credit assessment.