基于多角度对比学习的属性网络异常检测

Anomaly Detection on Attribute Network by Multi-Angle Contrastive Learning

传统属性网络的异常检测多基于自监督对比学习,存在着异常节点类型较少、对比学习角度单一等不足。基于“节点-节点”对比和“节点-子图”对比,提出了一种多角度对比学习属性网络异常检测的ADMC(anomaly detection on attribute network by multi-angle contrastive learning)模型。主要创新工作有:在原有结构和属性异常的基础上,进一步细化为四种节点异常类型,并在属性网络中对其进行量化测度;利用“节点-节点”的节点级对比获得二重联接信息,利用“节点-子图”的跨节点级对比获得局部连接信息,并构建了二者互为补充的多角度对比学习模型。基于社交媒体及引文文献等数据集,与基于自监督对比学习(CoLA)模型相比较的实验结果表明,ADMC模型能够丰富异常数据的类型,提升属性网络异常检测的精确性。

Traditional anomaly detection in attribute networks largely relies on self-supervised contrastive learning,which often falls short due to the limited types of anomalous nodes and the singularity in contrastive learning perspectives.This paper proposes an ADMC(anomaly detection on attribute network by multi-angle contrastive learning)model based on“node-node”contrast and“node-subgraph”contrast,aiming to address these deficiencies.The primary innovations include:refining the original structural and attribute anomalies into four distinct node anomaly types and quantifying them within the attribute network;employing“node-node”node-level contrasts to acquire biconnected information,and utilizing“node-subgraph”cross-node level contrasts to obtain local connectivity information,thereby constructing a multi-angle contrastive learning model where both perspectives complement each other.Through experiments on datasets from social media and citation literature,compared with the self-supervised contrastive learning(CoLA)model,the results demonstrate that the ADMC model can enrich the types of anomaly data and enhance the precision of anomaly detection in attribute networks.