Bayesian nonparametric dynamical clustering of time series

• Autores: Adrián Pérez Herrero
• Directores de la Tesis: Paulo Félix Lamas (dir. tes.) , Jesús María Rodríguez Presedo (dir. tes.)
• Lectura: En la Universidade de Santiago de Compostela ( España ) en 2026
• Idioma: inglés
• Programa de doctorado: Programa de Doctorado en Investigación en Tecnologías de la Información por la Universidad de A Coruña y la Universidad de Santiago de Compostela
• Enlaces
  • Tesis en acceso abierto en: MINERVA
• Resumen

  • Real-world systems often produce time series with recurring yet evolving patterns. Examples include the heartbeat’s cycle, waves reaching the shore, or machines switching between operating modes. Finding structure in this kind of data is still a major challenge in engineering and science. When the goal is to group segments with different shapes, traditional clustering methods run into three main difficulties. First, the segments are not independent; they occur in a temporal sequence that carries meaning. Second, the same underlying behaviour may appear with temporal distortions: local shifts, elongations, or compressions that must be accounted for. Third, the types of behaviours can change over time, with new patterns appearing and existing ones evolving. Balancing the need to remember known patterns while adapting to new ones is well-known as the stability-plasticity dilemma. This thesis aims to address these challenges in a principled and domain-agnostic way.