Anomaly detection on attributed networks is widely used in online shopping, financial transactions, communication networks, and so on. However, most existing works trying to detect anomalies on attributed networks only consider a single kind of interaction, so they cannot deal with various kinds of interactions on multi-view attributed networks. It remains a challenging task to jointly consider all different kinds of interactions and detect anomalous instances on multi-view attributed networks. In this paper, we propose a graph convolution-based framework, named AnomMAN, to detect Anomaly on Multi-view Attributed Networks. To jointly consider attributes and all kinds of interactions on multi-view attributed networks, we use the attention mechanism to define the importance of all views in networks. Since the low-pass characteristic of graph convolution operation filters out most high-frequency signals (aonmaly signals), it cannot be directly applied to anomaly detection tasks. AnomMAN introduces the graph auto-encoder module to turn the disadvantage of low-pass features into an advantage. According to experiments on real-world datasets, AnomMAN outperforms the state-of-the-art models and two variants of our proposed model.

翻译：属性网络上的异常检测广泛应用于在线购物、金融交易、通信网络等领域。然而，现有大多数针对属性网络的异常检测工作仅考虑单一类型的交互，无法应对多视图属性网络中的多种交互类型。如何联合考虑所有不同类型的交互并在多视图属性网络上检测异常实例仍是一项具有挑战性的任务。本文提出一种基于图卷积的框架——AnomMAN，用于多视图属性网络上的异常检测。为联合考虑属性及多视图属性网络中的所有交互类型，我们采用注意力机制定义网络中所有视图的重要性。由于图卷积操作的低通特性会滤除大部分高频信号（即异常信号），该方法无法直接应用于异常检测任务。AnomMAN引入图自编码器模块，将低通特征这一劣势转化为优势。在真实数据集上的实验表明，AnomMAN的性能优于当前最先进的模型及我们提出的两种变体模型。