Abstract
In the context of inductive learning, the Bayesian approach turned out to be very successful in estimating probabilities of events when there are only a few learning examples. The m-probability estimate was developed to handle such situations. In this paper we present the m-distribution estimate, an extension to the m-probability estimate which, besides the estimation of probabilities, covers also the estimation of probability distributions. We focus on its application in the construction of regression trees. The theoretical results were incorporated into a system for automatic induction of regression trees. The results of applying the upgraded system to several domains are presented and compared to previous results.