On open questions in the geometric approach to structural learning Bayesian nets (Q541844)

This paper presents a study on learning Bayesian network (BN) structures from data using the maximization-of-quality criterion. In the BN structure given by an acyclic directed graph \(G\), a certain vector with integers as components is applied. Every reasonable criterion in the form of a linear function plus a constant may be used to learn the BN structure. In this paper, the authors introduce a class of necessary linear constraints on a particular uniquely determined vector, here called the standard imset. Next, they formulate the conjecture that these constraints characterize the polytope \(P\). Thirdly, a previous hypothesis by Raymond Hemmecke that the only vectors with integers as components within \(P\) are standard imsets is confirmed. Three open (mathematical) questions related to this are analyzed in detail and proved. It is my belief that the presented proofs and results are valuable and worthy of notice. Also, the use of the linear transformation \(L\) introduced here makes it possible to simplify the algebraic representation of the BN structure.