DAMDID 2023. Keynote: Mikhail Zymber. Discord discovery it time series
Discord discovery in time series, or Can we detect all anomalies of an anomalously long time series in an anomalously short time?
A time series is a chronologically ordered real-valued sequence that reflects a certain process or phenomenon. Currently, time series are ubiquitous, and we need to store and process such data in a wide range of domains: the digital industry, personal healthcare, the Internet of Things, climate modeling, etc. In the above areas, discovering anomalies in time series remains one of the most topical problems. In addition, discovering subsequence anomalies is more challenging than detecting point outliers since a subsequence anomaly refers to successive points in time that are collectively abnormal, although each point is not necessarily an outlier. Since in the domains above, very long time series are typical, the discovery of subsequence anomalies requires parallel algorithms and high-performance computing. In the plenary talk, we are going to present the parallel subsequence anomaly discovery algorithms for GPUs and multi-GPU clusters that are developed in the Big Data and Machine Learning Lab of South Ural State University, Chelyabinsk, Russia.