融合双重自监督信号的图异常检测

图异常检测是网络研究中的一项重要内容。为解决以往工作中常依赖单一自监督信号而不能很好地检测多类型异常的问题，提出一种融合结构和属性的自监督图异常检测模型。首先选取目标节点，再基于图元邻接矩阵采样得到对应的负例节点；其次，构造正负子结构，并基于图卷积网络学习子结构表示以得到结构自监督信号；再次，依托自编码器对属性进行重构以获得属性自监督信号，解决节点匿名化带来的属性平滑问题；最后，通过对比学习对重构前后的正负实例对进行差值学习，以实现异常检测。在4个数据集上进行了3组实验，结果表明模型能够有效检测图中的异常节点。

Graph Anomaly Detection Integrating Dual Self-supervised Signals

Graph anomaly detection is an important content in network research. In order to solve the problem that the previous work often relies on a single self-supervised signal and cannot detect multiple types of anomalies well, the integrating structure and attribute self-supervised graph anomaly detection model is proposed. Firstly, the target node is selected, and then the corresponding negative node is sampled based on the adjacency matrix of graphlet. Secondly, the positive and negative substructures are constructed, and the substructure representation is learned based on graph convolutional network to construct the structure self-supervised signal. Thirdly, the attribute self-supervised is reconstructed by autoencoder to obtain the attribute self-supervised signal, to solve the attribute smoothing problem caused by node anonymity. Finally, the difference is learned by contractive learning between positive and negative pairs before and after reconstruction to realize anomaly detection. Three sets of experiments are carried out on four data sets, and the results show that the model can detect abnormal nodes in the graph effectively.