Ensemble techniques combine several individual classifiers to obtain a composite classifier that outperforms each of them alone. Despite of these techniques have been successfully applied to many domains, their applications on networked data still need investigation. There are not many known strategies for sampling with replacement from interconnected relational data. To contribute in this direction, we propose a cross-validated committee ensemble procedure applied to graph-based classifiers. The proposed ensemble either maintains or significantly improves the accuracy of the tested relational graph-based classifiers. We also investigate the role played by diversity among the several individual classifiers, i.e., how much they agree in their predictions, to explain the technique success or failure. © 2012 IEEE.