Loan Repayment Willingness Prediction Based on Graph Neural Networks

本文提出了一种基于图神经网络(Graph Neural Networks, GNNs)的方法用于预测欠款人的还款意愿。还款意愿是信贷风险管理中的关键因素，准确预测还款意愿对于降低违约风险具有重要意义。传统方法如评分卡模型虽然在信贷评估中被广泛应用，但其线性假设和对复杂关系建模能力的不足限制了预测精度。图神经网络能够有效处理数据中的复杂关系和非线性特征，通过构建借款人之间的关系图，并利用图神经网络学习节点(借款人)的嵌入表示，从而更准确地捕捉还款意愿的影响因素。实验结果表明，与传统评分卡模型相比，图神经网络方法在预测准确率、AUC值和F1分数等指标上均表现出显著优势，全面优于评分卡模型，为信贷风险评估提供了一种更高效、更精准的新方法。This paper proposes a method based on Graph Neural Networks (GNNs) to predict the repayment willingness of borrowers. Repayment willingness is a key factor in credit risk management, and accurately predicting it is of great significance for reducing the risk of default. Traditional methods, such as credit scoring card models, are widely used in credit assessment. However, their linear assumptions and limitations in modeling complex relationships restrict the accuracy of predictions. Graph Neural Networks are capable of effectively handling complex relationships and non-linear features in the data. By constructing a relationship graph among borrowers and using GNNs to learn the embedded representations of nodes (borrowers), the factors influencing repayment willingness can be more accurately captured. Experimental results show that, compared with traditional credit scoring card models, the GNN method demonstrates significant advantages in terms of prediction accuracy, AUC value, and F1 score. It outperforms the scoring card models comprehensively, providing a more efficient and precise new approach for credit risk assessment.