Skip to main content Skip to main navigation

Publication

Which Time Series Domain Shifts can Neural Networks Adapt to?

Henri Hoyez; Bruno Mirbach; Jason Raphael Rambach; Cedric Schockaert; Didier Stricker
In: Proceedings of. European Signal Processing Conference (EUSIPCO-2024), IEEE, 2024.

Abstract

Machine Learning on time series has great potential in solving real-life problems, from medical monitoring to machine predictive maintenance. In practice, however, the deployment domain can differ from the training conditions. These domain shifts give rise to the domain adaptation field which extends to time series data. However, despite the fact that broad domain shift definitions were already proposed, there is no clear definition adapted specifically to time series in the current literature. Moreover, even though evaluation methods were proposed, evaluations on real data do not allow a full control over the benchmarks. In this paper, we first propose a novel definition of domain shift based on a State Space Model (SSM). Then, we introduce a new dataset based on these SSMs to provide a benchmark using controlled domain shifts. Lastly, we perform an evaluation of the State of the Art on these generated domain adaptation problems, and hence systematically evaluate the domain shift effects on the adaptation performances.