Data mining and machine learning of multivariate time series is a difficult task. Because of varying lengths of time series, varying development over time, or possible long term correlations traditional machine learning algorithms do not qualify for this type of data.

Data mining and machine learning of multivariate time series is a difficult task. Because of varying lengths of time series, varying development over time, or possible long term correlations traditional machine learning algorithms do not qualify for this type of data. Embedding time series in a consistent feature space is necessary, which mostly conflicts with the varying structure of the data, and solutions like truncating or time series features are often impractical or come along with loss of information. Other approaches, e.g. alignment-kernel, may be successful on small data sets but are time-consuming and therefore problematic. Ordered means models present excellent starting points for this project. Machine learning of multivariate time series in a highly dynamic environment requires high degrees of adaptation and online learning capability. This project addresses exactly this problem, namely, to explore how OMMs can be extended such that the algorithms will be enabled to adapt learned models to new and unknown data and, thus, to new and unknown situations. Further concerns in dynamic environments are the tasks of unsupervised learning. Latest developments do not only learn and analyze sequences, but also generate new instances of learned time series data.

We develop models for unsupervised and reinforcement learning of hierarchical structure from sequential data both for the analysis and generation of multivariate sequence patterns, in particular, unsuspervised hierachical approaches.