This paper investigates an extension of classification trees to deal with uncertain information where uncertainty is encoded in possibility theory framework. Class labels in data sets are no longer singletons but are given in the form of possibility distributions. Such situation may occur in many real-world problems and cannot be dealt with standard decision trees. We propose a new method for assessing the impurity of a set of possibility distributions representing instances’s classes belonging to a given training partition. The proposed approach takes into account the mean similarity degree of each set of possibility distributions representing a given training partition. The so-called information closeness index is used to evaluate this similarity. Experimental results show good performance on well-known benchmarks.