The second factor explaining the poor performance of conventional PETs is that, when a
purely frequency-based (unsmoothed) estimate is used, small leaves give poor probability
estimates. This is the probability-estimation counterpart of the well-known “small disjuncts
problem”: in induced disjunctive class descriptions, small disjuncts are more error-prone
(Holte, Acker, & Porter, 1989). While this is not surprising statistically, the uniformity
and magnitude of the improvement given by the simple, easy-to-use, Laplace correction
nevertheless is remarkable.