Using node ordering to improve Structure MCMC for Bayesian Model Averaging
Abhineet Sharma · 2016
In this thesis, I address an important problem of estimating the structure of Bayesian network models using Bayesian model averaging approach. Bayesian networks are probabilistic graphical models which are widely used for probabilistic inference and causal modeling. Learning the structure of Bayesian networks can reveal insights into the causal structure of the underlying domain. Owing to the super exponential structure space, it is a challenging task to find the most suitable network model that explains the data. The problem is worsened when the amount of available data is modest, as there might be numerous models with non negligible posterior. Therefore, we are interested in the calculation of posterior of a feature like presence of an edge from one particular node to another or a particular set being a parent of a specific node. The contribution of this thesis includes a Markov Chain Monte Carlo simulation approach to sample network structures from a posterior and then using Bayesian model averaging approach to estimate the posterior of various features.