High-performance anomaly discovery in time series

Summary

People like to measure everything over time. Thus, time series are ubiquitous in a wide spectrum of real-life domains, including economics and finance, climate and weather, medicine and personal healthcare, digital industry and IoT, etc. Anomaly discovery has become the hot topic in time series mining. The reason is that people want to know ranges in time series that are the most dissimilar to all the rest. No doubt, people want to discover the ranges above as fast as possible. In this tutorial, we will present our work on parallel algorithms for time series anomaly discovery on GPUs. Using our developments, the tutorial participants will solve some matter-relevant real-life problems. A participant is not required to be a professional application programmer, since we will prepare datasets, program skeletons, and hints to turn the solution into a dozen lines of code.

Audience & prerequisites

Bachelor, Master, and PhD students, as well as researchers and practitioners (in Computer Science or close domains), are welcome. Experience in programming with any high-level programming language is required (Python is preferable). A participant is expected to have a valid Google account with access to Google Colab, as well as bring his/her laptop with a stable Internet connection.

Materials for the tutorial.

Timing and syllabus

The tutorial lasts four academic hours (with a five-minute break after each hour), where theory-hearing and practice-programming activities alternate. The tutorial covers the following topics: time series discords [1]; DRAG [2] and MERLIN [3], serial discord discovery algorithms; PD3 [4] and PALMAD [5], parallel discord discovery algorithms.

Lecturer and Instructor

Mikhail Zymbler

Doctor of Science (Physics and Mathematics)
South Ural State University, Chelyabinsk, Russia
Deputy Director of the Scientific and Educational Center “Artificial Intelligence and Quantum Technologies”

Scopus ID 55841425200

WoS Researcher ID L-2224-2013

 

 

Yana Kraeva

Candidate of Science (Physics and Mathematics)
South Ural State University, Chelyabinsk, Russia
Head of the Data Mining Department of the Scientific and Educational Center “Artificial Intelligence and Quantum Technologies”

Scopus ID 57205379857

WoS Researcher ID LKK-6818-2024

 

 

References

1. Lin J., Keogh E.J., Fu A.W., Herle H.V. Approximations to magic: Finding unusual medical time series. 18th IEEE Symp. on Computer Based Medical Systems (CBMS 2005), 23-24 June 2005, Dublin, Ireland. 2005. P. 329-334. DOI: 10.1109/CBMS.2005.34.
2. Yankov D., Keogh E.J., Rebbapragada U. Disk aware discord discovery: Finding unusual time series in terabyte sized datasets. Knowl. Inf. Syst. 2008. Vol. 17, no. 2. P. 241-262. DOI: 10.1007/s10115-008-0131-9.
3. Nakamura T., Mercer R., Imamura M., Keogh E.J. MERLIN++: Parameter-free discovery of time series anomalies. Data Min. Knowl. Discov. 2023. Vol. 37, no. 2. P. 670-709. DOI: 10.1007/s10618-022-00876-7.
4. Kraeva Ya., Zymbler M. A Parallel Discord Discovery Algorithm for a Graphics Processor. Pattern Recognition and Image Analysis. 2023. Vol. 33, no. 2. P. 101-112. DOI: 10.1134/S1054661823020062.
5. Zymbler M., Kraeva Ya. High-Performance Time Series Anomaly Discovery on Graphics Processors. Mathematics. 2023. Vol. 11, no. 14. Article 3193. DOI: 10.3390/math11143193.