On the robustness of Bayesian networks to learning from non-conjugate sampling