This paper from Pérez-Herrero and colleagues introduces a Bayesian nonparametric approach to dynamical clustering of time series. The model treats observations as generated by an unbounded number of clusters, each with linear dynamics, and allows the system to switch among an unknown number of regimes over time. A hierarchical Dirichlet process serves as the prior over the switching linear dynamical system parameters, while Gaussian process priors capture amplitude and temporal-alignment variation within clusters.
Because the number of clusters and regimes is not fixed in advance, the method can adapt to data without overfitting. The authors argue that explicitly modeling pattern evolution avoids unnecessary cluster proliferation. Inference is done through a variational lower bound, with both offline and online formulations, making the approach suitable for streaming settings.
The method is demonstrated on electrocardiogram analysis using publicly available databases. The arXiv version is a revised manuscript with methodological refinements, expanded experimental results, and improved presentation. As the only source reviewed here, no independent comparison or disagreement is available.