Cross-validation relies on the numb

Cross-validation relies on the number of folds (K) to consider, we experimentally set this parameter. Using K = 5 means that 5 different trainings are performed for each class of instances. For each training, one fold is considered as a test set and %-gaps are only computed on test folds to reflect the abilities of the new generated heuristics to solve new instances. To summarize, we have 9 classes of instances and each class experiences 5 trainings with different training and test sets. Therefore, we should obtain 5 heuristics per class of instances, i.e., one per fold. In the end, we only retained the best one among the 5 folds.