Uplift modeling aims to estimate the incremental impact of a treatment, such as a marketing campaign or a drug, on an individual's behavior. These approaches are very useful in several applications such as personalized medicine and advertising, as it allows targeting the specic proportion of a population on which the treatment will have the greatest impact. Uplift modeling is a challenging task because data are partially known (for an individual, responses to alternative treatments cannot be observed). In this paper, we present a new tree algorithm named UB-DT designed for uplift modeling. We propose a Bayesian evaluation criterion for uplift decision trees T by dening the posterior probability of T given uplift data. We transform the learning problem into an optimization one to search for the uplift tree model leading to the best evaluation of the criterion. A search algorithm is then presented as well as an extension for random forests. Large scale experiments on real and synthetic datasets show the eciency of our methods over other state-of-art uplift modeling approaches.