基于空间-邻域自适应的图卷积神经网络信贷欺诈检测模型  

A Graph Convolutional Credit Fraud Detection Model with Spatial-Neighborhood Adaptability

作　　者：张岚泽 赵晓亮 刘津彤 顾益军[1] Zhang Lanze;Zhao Xiaoliang;Liu Jintong;Gu Yijun(Department of Information and Network Security,People’s Public Security University of China,Beijing 100038,China;Beijing Chaoyang District People’s Court,Beijing 100021,China)

机构地区：[1]中国人民公安大学信息网络安全学院,北京100038 [2]北京市朝阳区人民法院,北京100021

出　　处：《数据分析与知识发现》2024年第4期137-151,共15页Data Analysis and Knowledge Discovery

基　　金：中国人民公安大学基本科研业务费项目(项目编号:2021JKF420)的研究成果之一

摘　　要：【目的】为信贷欺诈检测提供兼具空间和邻域自适应性的图卷积神经网络模型。【方法】提出双曲跳跃图卷积神经网络。在空间自适应方面,将节点属性表示为双曲空间可训练曲率,从而完成欺诈网络的低失真嵌入表示;在邻域自适应方面,定义双曲跳跃连接框架(HJK-Net)框架,通过双曲层间聚合机制对邻域表示结果进行融合。从而为关系网络提供融合空间和邻域自适应性的图表示学习结果,进而完成信贷欺诈检测任务。【结果】通过在公开且来源于实际业务场景的大型社交网络中部署实验,所提模型的AUC指标达到0.8335,相比于以GraphSAGE(NS)为代表的基线模型提升0.0594。【局限】浅层社交网络对邻域自适应性的优势略有限制,所提模型在大型复杂深度网络结构中优势更加明显。【结论】空间自适应为节点属性相关性提供更准确描述,邻域自适应为图表示学习选择最优的邻域聚合范围;融合空间和邻域自适应的模型在大型欺诈关系网中具备更好的识别效果。[Objective]This paper provides a graph convolutional neural network model with spatial and neighborhood adaptability for credit fraud detection.[Methods]We proposed a Hyperbolic Jumping Connection Graph Convolutional Neural Networks.Regarding spatial adaptability,we represented the node attributes as a trainable curvature in hyperbolic space and completed the low-distortion embedding representation of the fraudulent network.In terms of neighborhood adaptability,we defined a Hyperbolic Jumping Knowledge Networks framework and fused the neighborhood representation results through the hyperbolic inter-layer aggregation mechanism.As a result,we provided the relational network with a graph representation learning result integrating spatial and neighborhood adaptability.Finally,we completed the task of credit fraud detection.[Results]By deploying experiments in a large-scale social network that is publicly available and comes from real business scenarios,the proposed model achieved an AUC of 0.8355,which was 0.0594 higher than the baseline model represented by GraphSAGE(NS).[Limitations]The advantages of shallow social networks on neighborhood adaptability are slightly limited,and the advantages of our model are more evident in large-scale complex deep network structures.[Conclusions]Spatial adaptation provides a more accurate description of node attribute correlations,and neighborhood adaptation selects the optimal neighborhood aggregation range for graph representation learning.The proposed model has a better identification effect in large-scale fraud relationship networks.

关 键 词：图卷积神经网络 图表示学习 双曲空间 空间-邻域自适应性 信贷欺诈检测 

分 类 号：TP393[自动化与计算机技术—计算机应用技术] G250[自动化与计算机技术—计算机科学与技术]