A stratified k-fold partition with k = 10 is generally a good starting point. This is well explained on Wikipedia:

http://en.wikipedia.org/wiki/Cross-validation_(statistics)#K-fold_cross-validation

"Stratified" means that you should keep the original balance between positive record and negative records on every fold.

Dear Simone,

Actually, I am looking for a more intelligent subsample algorithm which is able to select representative samples and ignore outliers.