Several alternative approaches for learning probability trees have been proposed
in the literature but currently no thorough comparison of these approaches
exists. Hence, it is unclear which approaches are preferable under which circumstances.
The goal of this paper is to compare the main existing approaches and
a novel variant. We incorporated them in the relational decision tree learner
Tilde [2] and evaluate them by performing experiments on benchmark datasets
and on manipulated datasets. We use both non-relational and relational datasets.