Statistical modelling with additive Gaussian process priors

Ishida, Sahoko (2024) Statistical modelling with additive Gaussian process priors. PhD thesis, London School of Economics and Political Science.

Text - Submitted Version Download (5MB)

Abstract

Regression with Gaussian process (GP) priors has become increasingly popular due to its ability to model complex relationships between variables and handle auto-correlation in the data through the covariance function of the process, called kernel. Despite its popularity, the statistical modelling aspect of GP regression has received relatively limited attention. In this thesis, we explore a regression model where the regression function can be decomposed into a sum of lower-dimensional functions, akin to the principles of Generalised Additive Models (Hastie and Tibshirani, 1990). We propose additive interaction modelling using a class of hierarchical ANOVA decomposition kernel. This flexible statistical modelling framework naturally accommodates interaction effects of any order without increasing the number of model parameters. Our approach facilitates straightforward assessment and comparison of models with different interaction structures through the model marginal likelihood. We also demonstrate how this framework enhances the interpretability of complex data structures, especially when combined with the concept of kernel centring. The second segment of the thesis focuses on the computational aspects of implementing the proposed additive models for handling large-scale data structured in multidimensional grids. Such structured data often arise in scenarios involving multilevel repeated measurements, as commonly seen in spatio-temporal analysis or medical, behavioural, and psychological studies. Leveraging the Kronecker product structure within the covariance matrix, we reduce the time complexity to O(n3) and storage requirements to O(n2). We extend existing work in the GP literature to encompass all models under hierarchical ANOVA decomposition kernels. Additionally, we address issues related to incomplete grids and various missingness mechanisms. We illustrate the practical application of our proposed methodologies using both simulated and real-world spatio-temporal and longitudinal data.

Item Type:	Thesis (PhD)
Additional Information:	© 2024 Sahoko Ishida
Library of Congress subject classification:	Q Science > QA Mathematics
Sets:	Departments > Statistics
Supervisor:	Bergsma, Wicher
URI:	http://etheses.lse.ac.uk/id/eprint/4612

Actions (login required)

Record administration - authorised staff only