This is intentional. The model that is displayed is always that
generated from the entire training dataset. The testing options
(percentage split, cross-validation etc.) only effect the evaluation of
the algorithm i.e. the statistics that are computed.
Michael Ward wrote:
I have been using WEKA 3.2.3 and have noticed the following anomoly.
When using classifier J48 (and possibly others) with a percentage
split option the decision tree that is displayed is not consistent
with the percentage split. Rather, the displayed tree appears to be
one which is generated from the entire input data set (as opposed to
just that percentage allocated for training). Is this an intended
effect, or do I have a misunderstanding of what should be occuring
with the system ?