The Library
Partial correlation based penalty functions and prior distributions for Gaussian graphical models
Tools
Carter, Jack Storror (2021) Partial correlation based penalty functions and prior distributions for Gaussian graphical models. PhD thesis, University of Warwick.
|
PDF
WRAP_Theses_Carter_J_2021.pdf - Submitted Version - Requires a PDF viewer. Download (5Mb) | Preview |
Official URL: http://webcat.warwick.ac.uk/record=b3748414~S15
Abstract
Graphical models are a useful tool for encoding conditional independence relations. A common goal is to select the graphical model that best describes the conditional independence relationships between variables given observations of these variables. Under the additional Gaussian assumption, conditional independence is equivalent to zero entries in the inverse covariance matrix Ɵ. Thus sparse estimation of Ɵ in turn specifies a graphical model and the associated conditional independencies. Popular frequentist methods for this often involve placing a penalty function on Ɵ and maximising a penalised likelihood, whilst Bayesian methods require specification
of a prior distribution on Ɵ.
Conditional independence relations are invariant to non-zero scalar multiplication of the variables, however in this thesis we show that essentially all current penalised likelihood methods and many prior distributions are not invariant to such transformations of the variables. In fact many methods are very sensitive to rescaling of the variables which can, and often does, result in a vastly different selected graphical model. To remedy this issue we introduce new classes of penalty functions and prior distributions which are based on partial correlations. We show that such penalty functions and prior distributions lead to scale invariant estimation and posterior inference on Ɵ.
We pay particular attention to two penalty functions in this class. The partial correlation graphical LASSO places an L1 penalty on the partial correlations whilst the spike and slab partial correlation graphical LASSO is a penalty function based on a spike and slab prior formulation. The performance of these penalty functions is compared to that of current popular penalty functions in simulated and real world settings. We also investigate spike and slab priors in general for Gaussian graphical models and point out that care must be taken when considering the positive definiteness of Ɵ. With this in mind we provide some theoretical results based on Wigner matrices.
Item Type: | Thesis (PhD) | ||||
---|---|---|---|---|---|
Subjects: | Q Science > QA Mathematics | ||||
Library of Congress Subject Headings (LCSH): | Gaussian processes -- Data processing, Graphical modeling (Statistics), Functions, Inverse, Analysis of covariance, Multivariate analysis, Statistics -- Graphic methods, Matrices | ||||
Official Date: | October 2021 | ||||
Dates: |
|
||||
Institution: | University of Warwick | ||||
Theses Department: | Department of Statistics | ||||
Thesis Type: | PhD | ||||
Publication Status: | Unpublished | ||||
Supervisor(s)/Advisor: | Smith, Jim ; Rossell, David | ||||
Extent: | x, 150 leaves : charts | ||||
Language: | eng |
Request changes or add full text files to a record
Repository staff actions (login required)
View Item |
Downloads
Downloads per month over past year