Room “Sala Seminari” - Abacus Building (ex U14)
Federated Learning of Probabilistic Graphical Models
Speaker: Pablo Torrijos Arenas,
Albacete Research Institute of Informatics, University of Castilla-La Mancha (Spain)
Abstract
Machine learning now informs decisions in medicine, finance, and public administration, where regulations such as the GDPR and the EU AI Act require that data stay where it was collected and that model decisions remain auditable.
Federated learning answers the first requirement, but it has focused almost entirely on deep neural networks, which are hard to interpret. Probabilistic Graphical Models are a natural fit for the second requirement, and yet they have been largely absent from the federated setting. In this talk we will explain why, and what can be done about it. We will cover three lines of work: structural consensus, where heterogeneous local Bayesian networks are merged under complexity bounds; federated structure learning, with FedGES for discrete data and a cumulant-based family of LiNGAM models for continuous variables that also supports exact unlearning; and federated classification with the FedAnDE family, including a model specifically designed for high dimensionality scenarios.
Short bio
Pablo Torrijos Arenas is completing his PhD in the Department of Computing Systems at the University of Castilla-La Mancha (Spain), where he also earned his BSc (2021) and MSc (2022) in Computer Engineering. His doctoral thesis, Federated Learning of Probabilistic Graphical Models, is going to be defended in July 2026. He is currently a visiting researcher at the MADLab of the University of Milano-Bicocca, hosted by Prof. Fabio Stella. His research focuses on distributed and federated learning of Probabilistic Graphical Models, in particular Bayesian networks and Bayesian network classifiers, and on their scalability to high-dimensional problems. His work has appeared in journals such as Machine Learning and Knowledge-Based Systems, and at conferences including AAAI, GECCO, and CEC.
Contact person for the seminar: fabio.stella@unimib.it