基于图卷积自编码器的多视图属性网络异常检测算法  

Detect anomalies on multi-view attributed networks based ongraph convolution autoencoder

作　　者：李贺[1] 彭以冲 张万园 马小科[1] 崔江涛[1] 黄健斌[1] He LI;Yichong PENG;Wanyuan ZHANG;Xiaoke MA;Jiangtao CUI;Jianbin HUANG(School of Computer Science and Technology,Xidian University,Xi'an 710126,China)

机构地区：[1]西安电子科技大学计算机科学与技术学院,西安710126

出　　处：《中国科学:信息科学》2025年第2期269-283,共15页Scientia Sinica(Informationis)

基　　金：科技部科技创新2030重大项目(批准号:2021ZD0201300)资助。

摘　　要：属性网络上的异常检测已经在垃圾邮件检测、金融欺诈检测和网络安全中的入侵检测等领域得到了广泛应用.然而,现有的研究大多将属性网络中不同类别的数据等同对待,无法精准检测出多类别的节点异常.为了解决这个问题,本文提出了一个基于图卷积自编码器的多视图属性网络异常检测框架AMEAN(anomaly detection on multi-view attribute networks).具体来说,AMEAN首先根据节点分类拆分多视图,再利用属性网络中丰富的语义信息,使用图卷积网络对每个视图进行处理.由于图卷积运算的低通特性过滤了大部分异常信号(高频信号),AMEAN还引入了自编码器模块.异常程度越大的节点,经过图卷积自编码器的重构误差越大.通过图卷积和自编码器的这种协同作用,AMEAN可以从结构和属性两个角度来衡量不同类别节点的异常程度.在合成数据集和真实世界数据集上的实验结果表明,AMEAN优于现有的模型,能精准检测出属性网络中的多类别异常节点.Anomaly detection on attributed networks has been widely applied in areas such as spam detection,financial fraud detection,and intrusion detection in cybersecurity.However,most existing studies treat different categories of data in attributed networks equally,failing to accurately detect multi-category node anomalies.To address this issue,this paper proposes a multi-view attributed network anomaly detection framework based on graph convolutional autoencoders,named AMEAN(anomaly detection on multi-view attribute networks).Specifically,AMEAN first splits the network into multiple views based on node classification and then uses graph convolutional networks to leverage the rich semantic information in the attributed network to process each view.Due to the low-pass filter nature of graph convolution operations,which filters out most anomaly signals(high-frequency signals),AMEAN also incorporates an autoencoder module.The greater the anomaly degree of a node,the larger the reconstruction error of the graph convolutional autoencoder.Through the synergistic effect of graph convolution and autoencoders,AMEAN can measure the anomaly degree of different categories of nodes from both structural and attribute perspectives.Experimental results on synthetic and real-world datasets show that AMEAN outperforms existing models,accurately detecting multi-category anomalous nodes in attributed networks.

关 键 词：多视图属性网络 异常检测 图卷积网络 自编码器 

分 类 号：TP311.13[自动化与计算机技术—计算机软件与理论] TP18[自动化与计算机技术—计算机科学与技术] O157.5[理学—数学]