Hybrid semiparametric Bayesian networks

Abstract This paper presents a new class of Bayesian networks called hybrid semiparametric networks, which can model data (discrete and continuous data) by mixing parametric nonparametric estimation models. The models represent conditional linear Gaussian relationship between variables, while the other types relationships, such as non-Gaussian nonlinear relationships. generalizes including them special case. In addition, we describe learning procedure for structure parameters our proposed type network. finds best combination automatically from data. requires definition cross-validated score. We also detail how be sampled network, in turn useful to solve related tasks, inference. Furthermore, intuitively relate proposal with adaptive kernel density experimental results show that are valuable contribution when dealing do not meet assumptions expected models, networks. include experiments synthetic real-world UCI repository demonstrate good performance ability extract information about variables model.