【作者】 王宇；

【导师】 张凤斌；

【作者基本信息】 哈尔滨理工大学 ， 软件工程（专业学位）， 2022， 硕士

【摘要】 属性网络的异常检测研究,旨在发现网络中属性和特征与大多数节点存在较大差异的异常节点。然而,现有大多数方法忽略了网络结构信息与节点属性信息之间的联合交互问题,且由于网络噪声的影响导致节点特征学习质量低下,影响了异常检测算法的性能。本文提出了一种深度联合表示学习框架,进而提出异常检测算法。为了捕捉属性网络中网络结构和网络节点属性信息之间的跨模态交互信息,并且得到高质量的低维特征表示,本文设计了关联对抗正则化联合学习模型(Association Adversarial Regularization Joint Learning Model,AARJL),使用两个编码器联合学习隐向量空间中网络结构特征嵌入和节点属性特征嵌入,通过对抗正则化模块对两种特征表示进行正则化,将两个特征进行融合,实现对网络结构和网络节点属性的跨模态联合学习,在融合特征上引入超球面学习机制,通过计算潜空间中特征融合节点到超球中心的距离来检测异常。为了解决图卷积神经网络在结构特征提取时,因网络堆叠层数过多,导致模型特征提取性能下降的问题,引入非平滑聚合的节点特征提取思想来缓解图卷积神经网络各层之间过度平滑的现象,并在AARJL的基础上提出了关联对抗正则化属性网络异常检测方法(Associative Adversarial Regularized Attributed Network Anomaly Detection,AARAN)。在图卷积网络提取网络结构特征信息时,应用特征聚合在网络中逐层提取语义信息,并通过语义对齐使语义信息和网络结构特征映射到相同邻域。本文方法在Cora、Citeseer、Pubmed三个数据集上对比目前最新的方法表明,相较于传统基于图的异常检测方法,AARAN在AUC分数上最高提升了4.62%,在AP分数上最高提升了5.02%,证明了AARAN异常检测方法提高了检测属性网络中异常信息的能力。更多还原

【Abstract】 The research of anomaly detection in attributed network aims to find the abnormal nodes which are different from most of the nodes in attribute and feature.However,most existing methods ignore the joint interaction between network structure and node attributed,and the network can only learn low quality node feature because of the network noise,which influences the effectiveness of anomaly detection algorithm.This dissertation proposes not only a deep joint representation learning framework but an anomaly detection algorithm based on the above framework.In order to capture the cross-model interaction between network structure,and node node attribute.This dissertation designed an Association Adversarial Regularization Joint Learning Model.The model uses two encoders to learn network structure feature embedding and node attribute feature embedding jointly in the hidden vector space,and regularizes the two feature representations by adversarial regularization module,then fuses the two features to realize the cross-model joint learning of network structure and network node attribute.Finally,a hyperspherical learning mechanism is introduced on the fusion features to detect anomalies by calculating the distance between feature fusion nodes and the center of the hypersphere in low-dimensional latent space.In order to solve the problem in which model has a descend in feature extraction performance due to the stacked of layers in the process of structural feature extraction,this dissertation has introduce the idea of non-smooth aggregation node feature extraction,which is introduced to alleviate the phenomenon of excessive smoothness between all layers of graph convolutional neural network.Based on AARJL,this dissertation proposed an Associative Adversarial Regularized Attributed Network Anomaly Detection(AARAN).Feature aggregation methods is used to extract semantic information layer by layer in graph convolutional network,and by the means of semantic alignment achieve mapping semantic information and network structural features to the same neighborhood.In this dissertation,put AARAN compared with the latest methods on Cora,Citeseer and Pubmed data sets.The results showed that AARAN improved the AUC score by 4.62% and AP score by 5.02% compared with the traditional graph-based anomaly detection method.It is proved that AARAN can improve the ability of anomaly detection in the attributed network.更多还原