Edinburgh Research Explorer

Improving the structure MCMC sampler for Bayesian networks by introducing a new edge reversal move

Research output: Contribution to journalArticle

  • Marco Grzegorczyk
  • Dirk Husmeier

Related Edinburgh Organisations

Original languageEnglish
Pages (from-to)265-305
Number of pages41
JournalMachine Learning
Volume71
Issue number2-3
DOIs
Publication statusPublished - Jun 2008

Abstract

Applications of Bayesian networks in systems biology are computationally demanding due to the large number of model parameters. Conventional MCMC schemes based on proposal moves in structure space tend to be too slow in mixing and convergence, and have recently been superseded by proposal moves in the space of node orders. A disadvantage of the latter approach is the intrinsic inability to specify the prior probability on network structures explicitly. The relative paucity of different experimental conditions in contemporary systems biology implies a strong influence of the prior probability on the posterior probability and, hence, the outcome of inference. Consequently, the paradigm of performing MCMC proposal moves in order rather than structure space is not entirely satisfactory. In the present article, we propose a new and more extensive edge reversal move in the original structure space, and we show that this significantly improves the convergence of the classical structure MCMC scheme.

    Research areas

  • Bayesian networks, structure learning, MCMC sampling, MARKOV-CHAINS

ID: 13984313