Abstract: To address the issues that in attribute graph anomaly detection, existing methods fail to fully understand and handle the complex relationships between nodes and the multi-level information propagation mechanisms, making it difficult to comprehensively capture the diverse influences of different nodes on target nodes. A attributed graph anomaly detection method based on Multi-level Con-textual Correlation Analysis(MCCA) was proposed. This method measured the contextual correlation differences between each node and its neighboring nodes through canonical correlation analysis, thereby detecting anomalous nodes. First, normal nodes typically exhibit strong correlations with their neighbors, while anomalous nodes show weaker correlations. Based on this observation, MCCA employed a weight-sharing Graph Neural Network(GNN) to capture node embeddings and their neighboring node embeddings.Second, a neighbor aggregation module with a hierarchical attention mechanism was designed for contextual modeling. By distinguishing between inter-hop and intra-hop neighbor embeddings, this module effectively reduces the interference of noisy nodes. Finally, a correlation analysis module was introduced to maximize the contextual correlations of normal nodes, and anomalous nodes are detected through node contextual correlation analysis. Experimental results show that the proposed method outperforms the CoLA model on the Citeseer dataset with a 2.22% increase in AUC, and outperforms the Dual-SVDAE model on the Weibo dataset with a 1.6% increase in AUC. These results demonstrate that the hierarchical attention mechanism can effectively improve the anomaly detection capability of attribute graphs.

Key words: Keywords: anomaly detection, graph neural networks, canonical correlation analysis, graph embedding

摘要： 针对在属性图异常检测中，现有方法未能充分理解和处理节点间复杂的相互关系以及多级信息传递机制，导致难以全面捕捉不同节点对目标节点的多样化影响，提出一种基于多级上下文关联性分析的属性图异常检测方法（MCCA）。通过典型相关性分析来度量每个节点与其相邻节点之间的上下文相关差异，并据此检测异常节点。首先，正常节点与其邻居节点之间的相关性通常较强，而异常节点则表现出较弱的相关性。基于这一现象，MCCA利用权值共享的图神经网络来捕获节点嵌入及其相邻节点的嵌入。其次，本文设计了一个具有分层注意力机制的邻居聚合模块，用于上下文建模。该模块通过区分跳转间和跳转内的邻居嵌入，有效降低了噪声节点的干扰。最后，引入相关性分析模块以最大化正常节点的上下文相关性，并通过节点上下文相关性分析来检测异常节点。实验结果表明，所提出的方法在Citeseer数据集上与模型CoLA相比，在AUC上提升了2.22个百分点；在Weibo数据集上与模型Dual-SVDAE相比，在AUC上提升了1.6个百分点。可见分层注意力机制可以有效的提升属性图的异常检测能力。

关键词: 关键词: 异常检测, 图神经网络, 典型相关分析, 图嵌入