A meta-Gaussian approach to learning non-Gaussian Bayesian network structure

Most existing approaches to learning the structure of Bayesian networks assume that all variables are discrete or that all variables are continuously normally distributed. We propose a meta-Gaussian approach that is appropriate for direct learning from general, continuous variables. We first transform the original variables into standard normal variables. Under the assumption that the transformed variables are multivariate normally distributed, we then make use of existing algorithms to learn the network structure in the transformed space, and then project the results back into the original space. Preliminary experimental results show that this approach can recover the network structure, provided that the variables of the network satisfy a fundamental monotonicity property