Bayesian network structure learning with a quotient normalized maximum likelihood criterion

Learning the dependency structure of a multivariate distribution from the observational data is an important task since it allows us to speculate about the underlying causal mechanisms that induce dependencies. This structure learning task can be approached as a model selection problem. Recent studies have revealed that the popular Bayesian model selection criterion is not satisfactory. We review some of the information theoretic alternatives and introduce a new one called a quotient normalized maximum likelihood criterion.