On Bayesian approaches for analyzing temporally dependent data
Thesis event information
Date and time of the thesis defence
Place of the thesis defence
Hall L2, Linnanmaa Campus
Topic of the dissertation
On Bayesian approaches for analyzing temporally dependent data
Doctoral candidate
Master of Science Olli Sarala
Faculty and unit
University of Oulu Graduate School, Faculty of Science, Research Unit of Mathematical Sciences
Subject of study
Statistics
Opponent
Professor Arto Klami, University of Helsinki
Second opponent
Assistant Professor Henri Pesonen, Tampere University
Custos
Professor Mikko J. Sillanpää, University of Oulu
New Bayesian approaches for analyzing temporally dependent data
Many datasets contain dependencies that can be modelled as time series. Typically, a time series consists of sequential measurements or observations from an individual or target. Hence, the observed data often exhibit inherent temporal dependence, where upcoming values in the series may depend on its previous values. This poses a challenge to the analysis of such data, as the number of possible dependencies to account for increases rapidly compared to independent observations. Various models and estimation methods have been proposed for time series analysis.
This thesis work presents three probabilistic Bayesian statistical methods for analyzing temporally dependent datasets. These are Gaussian mixture models for clustering time-series-type data, non-negative matrix factorization, and dynamic network inference using Gaussian Graphical models via a chain of sparse partial correlation graphs. The common theme between the methods is accounting for the temporal dependency at a prior level, making it possible to treat the observations as conditionally independent. Performance of the proposed methods has been illustrated using datasets from, for example, biology and materials science.
This thesis work presents three probabilistic Bayesian statistical methods for analyzing temporally dependent datasets. These are Gaussian mixture models for clustering time-series-type data, non-negative matrix factorization, and dynamic network inference using Gaussian Graphical models via a chain of sparse partial correlation graphs. The common theme between the methods is accounting for the temporal dependency at a prior level, making it possible to treat the observations as conditionally independent. Performance of the proposed methods has been illustrated using datasets from, for example, biology and materials science.
Created 17.11.2026 | Updated 18.11.2026