Learning Node Abnormality with Weak Supervision
Learning Node Abnormality with Weak Supervision
复制标题
DOI：
10.1145/3583780.3614950
复制
发表时间：
2023-10
期刊：
影响因子：
--
通讯作者：
Qinghai Zhou;Kaize Ding;Huan Liu;H. Tong
中科院分区：
文献类型：
--
作者：
Qinghai Zhou;Kaize Ding;Huan Liu;H. Tong
图异常检测旨在识别非典型子结构，由于其对社交网络分析、安全、金融等各种应用领域的深远影响而吸引了越来越多的研究关注。缺乏对真实异常的先验知识一直是获取细粒度注释（例如异常节点）的主要障碍，因此，已经开发出大量现有方法，或者以有限数量的节点级监督或以无监督的方式。尽管如此，粗粒度图元素（例如，可疑的节点组）的注释通常需要很少的人力时间和专业知识，但相对更容易获得。因此，在弱监督环境中研究异常检测并建立不同粒度级别的注释之间的内在关系是很有吸引力的。在本文中，我们解决了粗粒度监督的弱监督图异常检测的挑战性问题，方法是：（1）提出一种新颖的图神经网络架构，具有名为 WEDGE 的注意机制，可以在给定异常子图的几个标签的情况下识别关键节点级异常；（2）设计一个具有对比损失的新目标，通过强制正常和异常图元素之间的独特表示来促进节点表示学习。通过对真实世界数据集的广泛评估，我们证实了我们提出的方法的有效性，与最佳竞争对手相比，AUC-ROC 提高了 16.48%。
Graph anomaly detection aims to identify the atypical substructures and has attracted an increasing amount of research attention due to its profound impacts in a variety of application domains, including social network analysis, security, finance, and many more. The lack of prior knowledge of the ground-truth anomaly has been a major obstacle in acquiring fine-grained annotations (e.g., anomalous nodes), therefore, a plethora of existing methods have been developed either with a limited number of node-level supervision or in an unsupervised manner. Nonetheless, annotations for coarse-grained graph elements (e.g., a suspicious group of nodes), which often require marginal human effort in terms of time and expertise, are comparatively easier to obtain. Therefore, it is appealing to investigate anomaly detection in a weakly-supervised setting and to establish the intrinsic relationship between annotations at different levels of granularity. In this paper, we tackle the challenging problem of weakly-supervised graph anomaly detection with coarse-grained supervision by (1) proposing a novel architecture of graph neural network with attention mechanism named WEDGE that can identify the critical node-level anomaly given a few labels of anomalous subgraphs, and (2) designing a novel objective with contrastive loss that facilitates node representation learning by enforcing distinctive representations between normal and abnormal graph elements. Through extensive evaluations on real-world datasets, we corroborate the efficacy of our proposed method, improving AUC-ROC by up to 16.48% compared to the best competitor.