A comparison of approaches for learning first-order logical probability estimation trees

Probability Estimation Trees (PETs) [9] try to estimate the probability with which an instance belongs to a certain class, rather than just predicting its most likely class. Several approaches for learning PETs have been proposed, mainly in a propositional context. Since we are interested in applying PETs in a relational context, we make some simple modifications to the first-order tree learner Tilde to incorporate the main approaches (and a novel variant) and we experiment with all of them. Our results provide insight into the strengths and weaknesses of the alternative approaches in a relational context.