融合知识的多视图属性网络异常检测模型  

Multi-view Outlier Detection for Attributed Network Based on Knowledge Fusion

作　　者：杜航原[1] 曹振武 王文剑[1,2] 白亮 DU Hang-Yuan;CAO Zhen-Wu;WANG Wen-Jian;BAI Liang(School of Computer and Information Technology,Shanxi University,Taiyuan 030006;Institute of Intelligent Information Processing,Shanxi University,Taiyuan 030006)

机构地区：[1]山西大学计算机与信息技术学院,太原030006 [2]山西大学智能信息处理研究所,太原030006

出　　处：《自动化学报》2023年第8期1732-1744,共13页Acta Automatica Sinica

基　　金：国家自然科学基金(U21A20513,62076154,61902227,62022052,62276159);山西省重点研发计划项目(202202020101003)资助。

摘　　要：属性网络异常检测在网络安全、电子商务和金融交易等领域中具有重要的理论与现实意义,近年来受到了越来越多的关注.大多数异常检测方法凭借网络有限的属性或结构信息进行决策生成,往往难以对异常模式做出可靠的描述.此外,网络节点对应的实体往往关联着丰富的领域知识,这些知识对于异常的识别具有重要的潜在价值.针对上述情况,提出一种融合知识的多视图网络异常检测模型,在多视图学习模式下通过数据与知识的互补融合实现了对异常节点的有效识别.首先,使用TransR模型由领域知识图谱抽取知识向量表示,并借助输入网络的拓扑关系构造其孪生网络.接着,在多视图学习框架下构建属性编码器和知识编码器,分别将属性网络及其孪生网络嵌入到各自的表示空间,并聚合为统一网络表示.最后,综合不同维度上的重构误差进行节点异常分数评价,从而识别网络中的异常节点.在真实网络数据集上的对比实验表明,提出的模型能够实现对领域知识的有效融合,并获得优于基线方法的异常检测性能.Outlier detection on attributed networks is of important theoretical and practical significance in the network security,ecommerce,financial transaction and many other fields,and receives more and more attentions in recent years.Most existing outlier detection methods usually generate decisions by pattern mining on the network structure or node attributes.However,it is difficult to make a reliable description for abnormal objects by just relying on the limited attribute and structure information directly available from given network data.Furthermore,the nodes in networks are usually associated with abundant domain knowledge in the real world,which has great potential value for outlier detection.To this end,this paper proposes a multi-view network outlier detection model based on knowledge fusion,which identifies the abnormal pattern effectively by complementary fusion of network data and associated knowledge under the multi-view learning mode.Firstly,the model applies TransR to extract knowledge vector representation from domain knowledge graph,and constructs a twin network with the topology structure of the input network.Then,the attribute encoder and the knowledge encoder are constructed under the multiview learning framework to embed he attributed network and its twin network into their respective representation spaces separately.On this basis,the network embeddings in two views are integrated into a unified representation by the aggregator.Finally,the abnormal score of each node is evaluated by integrating the reconstruction errors in the two different dimensions,and the abnormal nodes in the network are then recognized.Extensive experiments on real network datasets demonstrate that the proposed model can realize effective fusion of domain knowledge and acquire better outlier detection performance than baseline approaches.

关 键 词：属性网络 异常检测 图神经网络 知识融合 多视图学习 

分 类 号：O157.5[理学—数学]