We have presented a novel algorithm for learning Bayesian network structures from data. Our
algorithm computes the exact posterior probability of a queried local subnetwork, e.g., a directed
edge between two nodes. A modified version of the algorithm finds a global network structurewhich maximizes the posterior probability. A major advantage of this method is that it explores
all possible structures, still running “only” in an exponential time with respect to the number of
network variables. We can expect that this feature makes it possible to successfully analyze cases
where inexact methods may fail. Although we provided some bits of evidence for this hypothesis,
more convincing validation would require a dedicated comparison study, not pursued in this paper.
Here we, instead, reported a set of experiments to illustrate the presented methods and to investigate
the actual speed of the implemented algorithms.