Subgroup discovery is concerned with finding subsets of a population whose class distribution is significantly different from the overall distribution. Previously subgroup discovery has been predominantly investigated under the propositional logic framework. This paper investigates multi-class subgroup discovery in an inductive logic programming setting, where subgroups are defined by conjunctions in first-order logic. We present a new weighted covering algorithm, inspired by the Aleph first-order rule learner, that uses seed examples in order to learn diverse, representative and highly predictive subgroups that capture interesting patterns across multiple classes. Our approach experimentally shows considerable and statistically significant improvement of predictive power, both in terms of accuracy and AUC, and theory construction time, by considering fewer hypotheses.