One data entry is is left out at a time and its value predicted after
complete neural network calculations determined by the ANALYSIS
keyword, i.e., after fit or pruning of input variables in the remaining
data set. This procedure is repeated for all data entries. The LOO
calculations, however, are quite time demanding, since one complete
should be done for each data entry.
The neural network program also performs embedded virtual LOO (see Tetko
et al, 1995) for random option inPARTITION.
Both approaches provide very similar results using fit option in
i.e. when validation of neural networks for the pre-set number of input
variables is used. However, the results of virtual LOO if the pruning
option is selected in ANALYSIS, only corresponds
to validation of the variables selected by the pruning and are not comparable
with those of the full LOO procedure.