Scalable and accurate online multivariate anomaly detection

Abstract

The continuous monitoring of dynamic processes generates vast amounts of streaming multivariate time series data. Detecting anomalies within them is crucial for real-time identification of significant events, such as environmental phenomena, security breaches, or system failures, which can critically impact sensitive applications. Despite significant advances in univariate time series anomaly detection, scalable and efficient solutions for online detection in multivariate streams remain underexplored. This challenge becomes increasingly prominent with the growing volume and complexity of multivariate time series data in streaming scenarios. In this paper, we provide the first structured survey primarily focused on scalable and online anomaly detection techniques for multivariate time series, offering a comprehensive taxonomy. Additionally, we introduce the Online Distributed Outlier Detection (2OD) methodology, a novel well-defined and repeatable process designed to benchmark the online and distributed execution of anomaly detection methods. Experimental results with both synthetic and real-world datasets, covering up to hundreds of millions of observations, demonstrate that a distributed approach can enable centralized algorithms to achieve significant computational efficiency gains, averaging tens and reaching up to hundreds in speedup, without compromising detection accuracy.