Bayesian structure learning in graphical models using sequential Monte Carlo
(English)Manuscript (preprint) (Other academic)
In this paper we present a family of algorithms, the junction tree expanders, for expanding junction trees in the sense that the number of nodes in the underlying decomposable graph is increased by one. The family of junction tree expanders is equipped with a number of theoretical results including a characterization stating that every junction tree and consequently every de- composable graph can be constructed by iteratively using a junction tree expander. Further, an important feature of a stochastic implementation of a junction tree expander is the Markovian property inherent to the tree propagation dynamics. Using this property, a sequential Monte Carlo algorithm for approximating a probability distribution defined on the space of decompos- able graphs is developed with the junction tree expander as a proposal kernel. Specifically, we apply the sequential Monte Carlo algorithm for structure learning in decomposable Gaussian graphical models where the target distribution is a junction tree posterior distribution. In this setting, posterior parametric inference on the underlying decomposable graph is a direct by- product of the suggested methodology; working with the G-Wishart family of conjugate priors, we derive a closed form expression for the Bayesian estimator of the precision matrix of Gaus- sian graphical models Markov with respect to a decomposable graph. Performance accuracy of the graph and parameter estimators are illustrated through a collection of numerical examples demonstrating the feasibility of the suggested approach in high-dimensional domains.
Structure learning, Bayesian statistics, Gaussian graphical models
Probability Theory and Statistics
Research subject Mathematics
IdentifiersURN: urn:nbn:se:kth:diva-180326OAI: oai:DiVA.org:kth-180326DiVA: diva2:892652
QC 201605242016-01-112016-01-112016-05-24Bibliographically approved