PUB - Publikationen an der Universität Bielefeld

Learning from time series is in demand in many domains including finance, medicine and industry. Recently, a novel method for time series classification was proposed based on the idea of training self-predictive models on the time series and classifying in the space of the learned model parameters - the model space. In this thesis, learning in the model space of neural networks is investigated and extended. First, an empirical investigation of time series classification and clustering in the model space is conducted. Based on experiments on numerous time series datasets, key aspects are identified and improvements proposed. Then, the underlying concept is extended to transfer learning for time series. A novel approach for unsupervised transfer learning using self-predictive modelling is proposed. Finally, a modular framework for modelling parameterized processes is defined. The proposed approaches are successfully validated on synthetic and real-world datasets.