E-Locus - Institutional Repository of the University of Crete

Home Collections School/Department School of Sciences and Engineering Department of Computer Science Post-graduate theses

Post-graduate theses

Search command : Author="Στεφανίδης" And Author="Κωνσταντίνος"

Current Record: 26 of 824

[Add to Basket]

Identifier

000456945

Title

Choose wisely: an extensive evaluation of model selection for anomaly detection in time series

Alternative Title

Επιλέξτε προσεκτικά: μια εκτενής αξιολόγηση επιλογής μοντέλων για την ανίχνευση ανωμαλιών σε χρονοσειρές

Author

Συλλιγάρδος, Εμμανουήλ Π.

Thesis advisor

Τραχανιάς, Πσναγιώτης

Reviewer

Κομοντάκης, Νικόλαος
Παλπάνας, Θέμης

Abstract

Anomaly detection is a fundamental task for time-series analysis with important implications for the downstream performance of many applications. Despite increasing academic interest and the large number of methods proposed in the literature, recent benchmark and evaluation studies demonstrated that no overall best anomaly detection methods exist when applied to very heterogeneous time series datasets. This lack of a universally superior method poses a significant challenge for practitioners who need to select the most appropriate technique for their specific datasets. To overcome this limitation, this thesis proposes a model selection pipeline that can automatically determine the best anomaly detection technique based on the characteristics of the time series data. By leveraging time series classification algorithms for model selection, the goal is to provide a scalable and viable solution to solve anomaly detection over highly diverse time series collected from various domains. Existing AutoML solutions are not directly applicable to time series anomaly detection, and no evaluation of time series-based approaches for model selection currently exists. Accordingly, we compare 16 different classifiers over 1800 time series, representing a diverse range of datasets. By comparing the performance of these classifiers, the study provides the first comprehensive evaluation of time series classification as a model selection approach for anomaly detection. The results demonstrate that model selection methods outperform individual anomaly detection methods while maintaining execution times in the same order of magnitude. This evaluation serves as a crucial first step in demonstrating the accuracy and efficiency of time series classification algorithms for anomaly detection, set- ting a strong baseline that can guide the model selection step in general AutoML pipelines. The findings of this evaluation have significant implications for the field of time series anomaly detection. The demonstrated superiority of model selection methods over individual anomaly detection techniques highlights the importance of selecting the most appropriate method based on time series characteristics. By capitalizing on the strengths of different anomaly detection methods, practitioners can enhance the overall performance of their anomaly detection systems. More- over, the evaluation serves as a benchmark for comparing and selecting time series classification algorithms for model selection purposes in anomaly detection tasks. This comprehensive study not only provides valuable insights into the effectiveness of various classifiers, but also establishes a foundation for further research and development in automated model selection approaches for time series anomaly detection. Ultimately, the proposed model selection method and the experimental evaluation contribute to advancing the state-of-the-art in time series analysis and enable more accurate and efficient anomaly detection in diverse application do- mains.

Language

English

Subject

AutoML

Time series analysis

Time series classification

Ανάλυση χροσειρών

Αυτόματη μηχανική μάθηση

Ταξινόμηση χρονοσειρών

Issue date

2023-07-21

Collection