Representation learning for time series
We investigate specific representations of time series for solving classification tasks arising at ThyssenKrupp. Among other representation methods, we analyze the efficiency of a new method proposed by Verdun Lunel [1], based on delay embeddings combined with the Wasserstein distances, which are numerical costs of an optimal transportation problem. Once the Wasserstein distances of each reconstructed time series are obtained, the classical multidimensional scaling (MDS) is implemented for further visualization and classification.
[1] Sjoerd Verduyn Lunel - Using Dynamics to Analyse Time Series; in Patterns of Dynamics, Berlin, July 2016; Springer Proceedings in Mathematics & Statistics, Volume 205, 2017
Principle Investigators: Pavel Gurevich, Hannes Stuke
Members: Yu He
Collaboration partner:
Paul Alexandru Bucur (ThyssenKrupp)