Polyhedral aspects of score equivalence in Bayesian network structure learning (Q2364495)

This paper is devoted to the problem of learning Bayesian network structure by means of integer linear programs. A score equivalence concept is introduced and the criteria of belonging to the linear space of score equivalent objectives is proved. A correspondence between score equivalent faces of the family-variable polytope and some faces of the characteristic-imset polytope is established. For the family-variable polytope the authors specify those faces which identified with empty graphs, full graphs, extreme supermodular functions, and uniform matroids. They also investigate their conjecture about sufficiency of score equivalent faces in the family-variable polytope.