基于图卷积网络的个人信用评估研究  

Research on Personal Credit Evaluation Based on Graph Convolutional Network

作　　者：唐灵慧 李林[1] 李丹 TANG Linghui;LI Lin;LI Dan(School of Business,University of Shanghai for Science and Technology,Shanghai 200093;Shanghai Ying Fan Technology Company,Shanghai 200093)

机构地区：[1]上海理工大学管理学院,上海200093 [2]上海颖幡技术有限公司,上海200093

出　　处：《计算机与数字工程》2024年第3期768-774,820,共8页Computer & Digital Engineering

摘　　要：传统的机器学习模型无法表现出个人信用评估问题中贷款者之间的高维邻居关系,通过构造图数据,利用图卷积网络进行分类预测的方式可以考虑贷款者在多个信息维度的相互联系。首先,利用递归特征消除(Recursive Feature Elimination,RFE)进行特征选择,筛选出贡献度最大的特征集。其次,利用随机森林(Random Forest,RF)计算出筛选后特征的重要性权重,同时将特征划分为类别特征和数值特征,根据特征类型并结合特征权重计算贷款者之间的距离,从而得到邻接矩阵。最后,将构造的图数据输入图卷积网络(Graph Convolutional Network,GCN)进行训练并预测结果。基于公开的德国个人信用数据集,通过两种评价指标对比了该模型与4篇近年研究结果,以及通过4种评价指标对比了该模型与3种基准模型。最终结果显示该模型的预测结果均要优于其他模型,能够有效进行个人信用贷款评估问题研究。In order to address the problem that traditional machine learning models cannot represent the high-dimensional neighbor relationships among lenders in the personal credit assessment problem,this paper proposes a graph convolutional net-work-based personal credit assessment model from a network science perspective,taking into account the multidimensional interre-lationships among lenders.To avoid the impact of feature data redundancy on the accuracy of the model,firstly,recursive feature elimination is used to filter out the feature set that contributes most to the personal credit assessment.Secondly,the importance weights of the filtered features are calculated using random forest,and the features are classified into category features and numeri-cal features.The distance between lenders is calculated based on the feature types and feature weights to obtain the adjacency matrix of the lender network.Finally,the constructed adjacency matrix with lender feature data is input into graph convolutional network for training and predicting the results.Based on the publicly available German personal credit dataset,the model is compared with the results of four recent studies through two evaluation metrics,as well as with three benchmark models through four evaluation met-rics.The experimental results show that the prediction results of this method are all better than other models and can perform person-al credit assessment more accurately.

关 键 词：个人信用评估 图卷积网络 特征选择 特征重要性 随机森林 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]