Graph sampler: a C language software for fully Bayesian analyses of Bayesian networks
Sagnik Datta, Ghislaine Gayraud, Éric Leclerc, Frédéric Y. Bois · arXiv (Cornell University) · 2015
Bayesian networks (BNs) are widely used graphical models usable to draw statistical inference about network structures. We present here Graph sampler a fast free C language software for structural inference on BNs. Graph sampler uses a fully Bayesian approach in which the marginal likelihood of the data (the data prior predictive distribution) and prior information about the network structure are considered. Inference on continuously valued BNs is treated as a Gaussian regression problem where the regressands are children nodes and the regressors their parents. Normal-gamma or Zellner likelihoods can then be used. For the discrete case a Dirichlet-multinomial model is available. A MetropolisHastings algorithm with fast jump kernel is used to sample graphs from their joint posterior distribution. The source C code distributed is very compact, fast, uses low memory and disk storage. We also reviewed the key features and working environment of other available software. As an illustration we compared the performance of our code to that of Structmcmc, a R package, using simulated data sets for graphs of 5 to 120 nodes. Our results show that Graph sampler is time ecient in identifying the underlying network and is time ecient. Several problems that arise when using Zellner priors are discussed. The relative weights of prior and likelihood are also discussed.