PhD defence: Multivariate Time Series Pattern Search

This PhD thesis focuses on the topic of (visual) pattern search in multivariate time series. On this topic, we developed accurate, efficient, and interpretable algorithms and designed tools for domain users. Traditional methods were designed primarily for univariate time series pattern search with relatively distinctive and unambiguous target patterns. They may not extend naturally to multivariate cases, and their performance may deteriorate significantly in the presence of distortions. If based on machine learning, conventional techniques become inefficient and uninterpretable and the retrieval accuracy may stagnate.

Because it is unlikely that a single tool can fit all use cases, we proposed a toolbox of multiple methods, including 1) a scalable, steerable, and interpretable hashing-based representation for pattern search, especially in very high-dimensional time series; 2) an efficient technique capturing various pattern distortions, especially time shifts between tracks; 3) an accuracy-centric, model-agnostic, machine learning-based framework that is simultaneously more accurate and more efficient than the prevailing machine-learning-based pattern search framework; and 4) an enhancement of user feedback for active-learning-based, feedback-driven pattern search striving for the highest possible retrieval accuracy.

All our proposed algorithms and tools work in and some even prefer multivariate cases. Extensive experiments verified the aforementioned benefits regarding accuracy, efficiency, and if necessary the steerability and interoperability of the proposed methods. Moreover, case studies and expert studies validated the usability of the user interfaces accompanying the proposed algorithms. Our tools are helping automotive calibration engineers trace events of interest and enable further domain-specific analysis. They are domain agnostic and applicable to use cases in other domains.