Research on Loan Default Risk Prediction Methods Based on Graph Neural Networks

Traditional statistical inference and machine learning models have been widely utilized in the task of predicting loan default risk. However, these approaches neglect the unique relationships between financial entities, leading to the loss of valuable information between entities and significantly impacting the accuracy of default risk prediction. To address these challenges, this study proposes a graph neural network model for predicting loan default risk: initially converting the raw tabular structure of loan records into a loan data graph; subsequently employing a graph sampling learning module to extract information from the loan data graph; and ultimately achieving precise prediction of default risk through a default prediction module. The proposed method in this paper offers a novel perspective for forecasting loan default, emphasizing the crucial importance of capturing relationships between financial entities. Furthermore, a novel node aggregation technique distinct from other graph neural networks is introduced in this study, capable of extracting more meaningful information from the loan data graph, thus enabling more accurate prediction of default risk. The efficacy of the proposed model is verified through extensive experimentation on actual datasets. Ablation studies demonstrate the effectiveness of the submodule, while baseline comparison experiments indicating the model's superior performance.