基于自监督的端到端图数据异常检测方法

Self-supervised end-to-end graph level anomaly detection

异常检测旨在发掘数据中异于寻常的模式,它在金融欺诈以及网络入侵检测等领域有着广泛的应用前景.本文主要研究了如何在结构复杂的图数据中进行异常检测,这涉及到挖掘异常的图结构信息以及节点属性信息.现有大部分工作通常采用一个两步的框架,即先对结构复杂的图数据进行表征学习生成图表征向量,然后再将该向量用于下游异常检测任务.由于分开训练的图表征学习任务与下游异常检测任务存在一定的语义鸿沟,这导致现有方法无法有效地挖掘出图中潜在的异常模式.因此,我们提出了一种基于自监督的端到端图数据异常检测框架SGAD,它可以有效地捕获图数据的语义信息并用于异常检测.具体来说, SGAD对无标签图数据进行了一系列变换用于构建自监督辅助任务,然后该自监督任务的输出结果可以直接用于异常检测.我们在多个公开数据集上进行了大量实验,实验结果表明本文提出的SGAD与现有方法相比获得了显著的效果提升.

Anomaly detection aims to identify unusual patterns deviating from the majorities,which is widely used in financial fraud detection,network intrusion detection,etc.This paper mainly focuses on anomaly detection of anomaly structural and node attribute information in graph-structured data.Most existing methods usually follow a two-step anomaly detection procedure.They first perform representation learning on the graph,and then the learned graph representations are fed into the downstream anomaly detection task.However,due to the separate training process of representation learning and anomaly detection,they cannot detect anomaly patterns effectively.Therefore,we propose a self-supervised graph-level anomaly detection(SGAD)framework,which can detect anomaly patterns in an end-to-end manner.Specifically,SGAD designs a self-supervised pretext task to perform representation learning,and then the output of this task can be used for explicit anomaly detection.We conduct extensive experiments on six public datasets,and the experimental results demonstrate that the proposed SGAD can achieve state-of-the-art performance on all the datasets.