In this paper, we are interested in the predictive and discriminant nature of features in supervised classification problems. We discuss the notions of prediction and discrimination and propose a hybrid approach combining supervised classifiers, model explanation, multicriteria decision making and pattern mining for identifying the most predictive and discriminant features in a dataset. The explanation of models learned by supervised classifiers produces rankings of features according to various performance measures. Based on that, multicriteria decision making and pattern mining methods are used to, respectively, select the most important features and interpret their role in terms of prediction and discrimination. Finally, we present and discuss two experiments on public datasets illustrating the potential of the approach.