For a presented case, a Bayesian network classifier in essence
computes a
posterior probability distribution over its class variable. Based
upon this
distribution, the classifier's classification function returns a
single,
determinate class value and thereby hides the uncertainty
involved. To provide
reliable decision support, however, the classifier should be able to
convey
indecisiveness if the posterior distribution computed for the case does
not
clearly favour one class value over another. In this paper we
present an
approach for this purpose, and introduce new measures to capture the
performance and practicability of such classifiers.
(full paper)