I experienced in a predictive model for a dataset I have from experiments on human subjects, that the error of the prediction in 10-fold cross validation is less than leave-one-subject-out (leave-one-person-out). I think leave-one-person out has advantage over the K-Fold CV since it takes intra-individual variability into account and thus has higher generalization power than K-Fold. However I am not sure if I am missing any advantage of K-Fold over leave-one-subject-out.
Asked
Active
Viewed 37 times
0
-
Speed may be an issue for large datasets. You can do your K-Folds on individuals to deal with the intra-individual issue – Henry Jul 13 '18 at 09:28
-
While the question is somewhat different the answers do seem to address the issue here (which is the criterion for closure as duplicate). – Glen_b Jul 14 '18 at 01:39