Thesis defense: 'Dynamic non-parametric Bayesian clustering of time series'

Many systems reveal their nature through repetition: a heart that beats, a wave reaching a shore, a machine cycling through its modes of operation. The fundamental question is not only what forms repeat but also how many there are, how they evolve, and how they relate in sequence. This thesis presents the Hierarchical Dirichlet Process - Gaussian Process Clustering (HDP-GPC), a non-parametric Bayesian model that infers an unlimited number of evolving group prototypes from sequential data, with calibrated uncertainty and explicit temporal alignment. Validation in ECG arrhythmia analysis, clinical differentiation between Takotsubo syndrome and myocardial infarction, and clustering of ocean wave spectra demonstrates that parsimonious and interpretable clustering can become a tool in itself for discovering underlying phenomena.

Supervisors: Paulo Félix Lamas and Jesús Rodríguez Presedo