On Model Selection, Bayesian Networks, and the Fisher Information Integral

Abstract. We study BIC-like model selection criteria and in particular, their refinements that include a constant term involving the Fisher information matrix. We observe that for complex Bayesian network models, the constant term is a negative number with a very large absolute value that dominates the other terms for small and moderate sample sizes. We show that including the constant term degrades model selection accuracy dramatically compared to the standard BIC criterion where the term is omitted. On the other hand, we demonstrate that exact formulas such as Bayes factors or the normalized maximum likelihood (NML), or their approximations that are not based on Taylor expansions, perform well. A conclusion is that in lack of an exact formula, one should use either BIC, which is a very rough approximation, or a very close approximation but not an approximation that is truncated after the constant term.