Anomaly detection is an interdisciplinary research area which attracts substantial attention both in statistics and in machine learning due to its critical role in a wide range of diverse applications, from cybersecurity to health monitoring. Most recently, spatio-temporal modeling through contrastive graph learning emerges as a promising alternative in scenarios, where there exists limited or even non-existent records of labeled anomalies and the data exhibit a sophisticated nonlinear dependence structure. However, prevailing graph contrastive learning methods are based on restrictive assumptions regarding selection of anomalousness thresholds, which limits their scope of usage. Motivated by threat detection in cyber-physical systems, we develop a new approach to anomaly detection in spatio-temporal data by fuzing the notion of self-supervised graph contrastive learning and extreme value theory. Our key idea is to bolster the anomaly detection performance of neural networks by leveraging the complementary insight on the intrinsic multi-scale data organization which contrastive learning can provide. We then show how our ideas at the interface of extreme value theory and graph contrastive learning can be combined to deliver more systematic and reliable detection of anomalies. We illustrate the utility of the new approach to detecting anomalies in cyber-physical Internet of Things, water distribution and military computer network systems.
Zhen et al. (Tue,) studied this question.