Title:Warping Resilient Scalable Anomaly Detection in Time Series

Abstract:Time series data is ubiquitous in the real-world problems across various domains including healthcare, social media, and crime surveillance. Detecting anomalies, or irregular and rare events, in time series data, can enable us to find abnormal events in any natural phenomena, which may require special treatment. Moreover, labeled instances of anomaly are hard to get in time series data. On the other hand, time series data, due to its nature, often exhibits localized expansions and compressions in the time dimension which is called warping. These two challenges make it hard to detect anomalies in time series as often such warpings could get detected as anomalies erroneously. Our objective is to build an anomaly detection model that is robust to such warping variations. In this paper, we propose a novel unsupervised time series anomaly detection method, WaRTEm-AD, that operates in two stages. Within the key stage of representation learning, we employ data augmentation through bespoke time series operators which are passed through a twin autoencoder architecture to learn warping-robust representations for time series data. Second, adaptations of state-of-the-art anomaly detection methods are employed on the learnt representations to identify anomalies. We will illustrate that WaRTEm-AD is designed to detect two types of time series anomalies: point and sequence anomalies. We compare WaRTEm-AD with the state-of-the-art baselines and establish the effectiveness of our method both in terms of anomaly detection performance and computational efficiency.