Beatrice Franzolini (King's College London)
will speak on
Multivariate random partition learning and prediction
Time: 3:00PM
Date: Thu 26th November 2026
Location: N0.20 - Science North
[map]
Abstract: A number of models (including mixture models, stochastic block models, and mixtures of experts) capture data complexity by positing a latent partition of the observed data points. However, standard formulations can be restrictive for modelling repeated measurements on the same subjects. Examples include multivariate panels, dynamic networks, and multi-view data where cluster membership may differ across views and evolve over time. Patients who group together on one set of biomarkers may split apart on another, and nodes in a network may move between communities through time. In this talk, I will address the following question: how can a generative model capture these differences, learn what different views or time points have in common, and use evolving clustering processes to predict future observations?
Treating the partition as random within a Bayesian framework helps control overfitting while preserving flexibility. The symmetries of its law govern how learning is structured and which information is used for inference and prediction. Well-established exchangeable random partitions, characterised through Kingman’s representation theorem, induce a single univariate random partition (even when exchangeability is relaxed to partial exchangeability), typically resulting in loss of information when they are applied to repeated measurements and multi-view data.
The core of the talk focuses on how to move from univariate to multivariate random partitions and on their probabilistic symmetries and statistical properties.
(This talk is part of the Statistics and Actuarial Science series.)
PDF notice
Return to all seminars