Real-world machine learning applications must be able to adapt to systematic changes in the data, e.g. a new subject or sensor displacement. This can be seen as a form of transfer learning, where the goal is to reuse the old (source) model by adapting the new (target) data. This is a challenging task, if no labels for the target data are available. Here, we propose to use the structure of the source and target data to find a transformation from the source to target space in an unsupervised manner. Our preliminary experiments on multivariate time series data show the feasibility of the approach, but also its limits.