Paul Wurth (Luxembourg)

Time series data are generally easy to obtain but often suffer from issues such as incomplete labeling, missing values, and privacy constraints. Transferring data from one domain to another using Machine Learning offers a promising solution to these challenges. This paper introduces a novel feed-forward multi-style transfer algorithm for time series. The proposed approach utilizes dual encoders to disentangle content and style from input sequences, which are then recombined by a decoder to generate sequences with the specified characteristics. Additionally, we propose a new metric to evaluate domain shifts and quantify implicit differences between datasets. Our method demonstrates robust transfer performance across diverse datasets, ranging from synthetic datasets to multivariate human activity recognition time series.

International audience