Cross-Domain Graph Anomaly Detection
Cross-Domain Graph Anomaly Detection
复制标题
DOI：
10.1109/tnnls.2021.3110982
复制
发表时间：
2021-10
影响因子：
10.4
通讯作者：
Kaize Ding;Kai Shu;Xuan Shan;Jundong Li;Huan Liu
中科院分区：
文献类型：
--
作者：
Kaize Ding;Kai Shu;Xuan Shan;Jundong Li;Huan Liu
由于在网络安全、金融和医疗保健等各种高影响领域的广泛应用，归因图的异常检测最近受到越来越多的研究关注。迄今为止，由于获取异常标签的成本昂贵，尤其是对于新形成的域，大多数现有工作主要以无监督的方式进行。如何利用标记属性图中的宝贵辅助信息来促进未标记属性图中的异常检测很少被研究。在这项研究中，我们的目标是通过域适应来解决跨域图异常检测的问题。然而，这项任务仍然很重要，主要是因为：1）数据异质性，包括属性图中的拓扑结构和节点属性，2）捕获目标域图上的不变和特定异常的复杂性。为了应对这些挑战，我们提出了一种新颖的框架 COMMANDER，用于属性图的跨域异常检测。具体来说，COMMANDER 首先通过图注意编码器将来自不同域的两个属性图压缩到低维空间。此外，我们利用域鉴别器和异常分类器来检测来自不同域的网络中出现的异常。为了进一步检测仅出现在目标网络中的异常，我们开发了属性解码器来为评估节点异常提供额外的信号。对各种现实世界跨域图数据集的广泛实验证明了我们方法的有效性。
Anomaly detection on attributed graphs has received increasing research attention lately due to the broad applications in various high-impact domains, such as cybersecurity, finance, and healthcare. Heretofore, most of the existing efforts are predominately performed in an unsupervised manner due to the expensive cost of acquiring anomaly labels, especially for newly formed domains. How to leverage the invaluable auxiliary information from a labeled attributed graph to facilitate the anomaly detection in the unlabeled attributed graph is seldom investigated. In this study, we aim to tackle the problem of cross-domain graph anomaly detection with domain adaptation. However, this task remains nontrivial mainly due to: 1) the data heterogeneity including both the topological structure and nodal attributes in an attributed graph and 2) the complexity of capturing both invariant and specific anomalies on the target domain graph. To tackle these challenges, we propose a novel framework COMMANDER for cross-domain anomaly detection on attributed graphs. Specifically, COMMANDER first compresses the two attributed graphs from different domains to low-dimensional space via a graph attentive encoder. In addition, we utilize a domain discriminator and an anomaly classifier to detect anomalies that appear across networks from different domains. In order to further detect the anomalies that merely appear in the target network, we develop an attribute decoder to provide additional signals for assessing node abnormality. Extensive experiments on various real-world cross-domain graph datasets demonstrate the efficacy of our approach.