Segmentation-Based Methods for Top-k Discords Detection in Static and Streaming Time Series Under Euclidean Distance

• Published in: Context-Aware Systems and Applications Vol. 409; pp. 147 - 163
• Main Authors: Thuy, Huynh Thi Thu, Anh, Duong Tuan, Chau, Vo Thi Ngoc
• Format: Book Chapter
• Language: English
• Published: Switzerland Springer International Publishing AG 2022; Springer International Publishing
• Series: Lecture Notes of the Institute for Computer Sciences, Social Informatics and Telecommunications Engineering
• Subjects: Anomaly detection; Clustering; discords; Segmentation; Streaming time series; Top
• ISBN: 3030931781; 9783030931780
• ISSN: 1867-8211; 1867-822X
• DOI: 10.1007/978-3-030-93179-7_12

Detecting top-k discords in time series is more useful than detecting the most unusual subsequence since the result is a more informative and complete set, rather than a single subsequence. The first challenge of this task is to determine the length of discords. Besides, detecting top-k discords in streaming time series poses another challenge that is fast response when new data points arrive at high speed. To handle these challenges, we propose two novel methods, TopK-EP-ALeader and TopK-EP-ALeader-S, which combine segmentation and clustering for detecting top-k discords in static and streaming time series, respectively. Moreover, a circular buffer is built to store the local segment of a streaming time series and calculate anomaly scores efficiently. Along with this circular buffer, a delayed update policy is defined for achieving instant responses to overcome the second challenge. The experiments on nine datasets in different application domains confirm the effectiveness and efficiency of our methods for top-k discord discovery in static and streaming time series.