A Bayesian approach for comparing cross-validated algorithms on multiple data sets