Error propagation in graphical models

Organisatsiooni nimi
Chair of Data Science
Kokkuvõte
The thesis focuses on the conditional independence (CI) relations in Probabilistic Graphical Models (PGMs). PGMs, including Bayesian Networks and Markov Networks, use graph separators to express CI relations [1]. Current inference systems are built on the premise that the CI relations used to build PGMs hold exactly, but in practice, it is often only possible to find CIs that hold approximately. This project aims to explore how errors in the initial set of CIs propagate to the CIs inferred from the graphical model. The first objective is to outline the basic concepts and results [2]. A more theory-oriented thesis will then focus on the underlying mathematical proofs (challenging), and a more practice-oriented thesis will experimentally test the hypothesis that the best known upper bounds for the error propagation are in fact optimal.

[1] Daphne Koller and Nir Friedman. Probabilistic Graphical Models - Principles and Techniques. MIT Press, 2009.
[2] Batya Kenig. Approximate implication for probabilistic graphical models. CoRR, abs/2310.13942, 2023.
Lõputöö kaitsmise aasta
2024-2025
Juhendaja
Miika Hannula
Suhtlemiskeel(ed)
inglise keel
Nõuded kandideerijale
Tase
Bakalaureus, Magister
Märksõnad

Kandideerimise kontakt

 
Nimi
Miika Hannula
Tel
E-mail
miika.hannula@ut.ee