Abstract
The advantage of using linear regression in the leaves of a regression tree is analysed in the paper. It is carried out how this modification affects the construction, pruning and interpretation of a regression tree. The modification is tested on artificial and real-life domains where its impact on classification error and stability of the induced trees is considered. The results show that the modification is beneficial, as it leads to smaller classification errors of induced regression trees. The Bayesian approach to estimation of class distributions is used in all experiments.