Title: Bayesian nonparametric multilayer clustering of longitudinal data
Authors: Beatrice Franzolini - Agency for Science, Technology and Research, Singapore (Singapore) [presenting]
Maria De Iorio - UCL (United Kingdom)
Abstract: A new class of Bayesian nonparametric models is introduced to make inferences on an ordered collection of partitions of the same objects, namely, a multilayer partition. The class is suited to analyze panel/longitudinal data, where repeated observations are collected over time for the same observational units. The core is a conditional partial exchangeable structure, which we argue is a natural and general modeling strategy for this context. The resulting class of Bayesian models guarantees analytical and computation tractability both in terms of the clustering structure and the underlying random probabilities measures. It allows predictions for any new number of observations and may, ultimately, constitute a powerful reference framework -currently missing in the literature- for the development of tailored Bayesian nonparametric models for panel data. We further explore in detail two specific models within this class: one based on a novel prior and another employing the well-known hierarchical Dirichlet process as a building block.