On May 1, 2005, at 12:46 AM, zia wrote:
I am confused with results....
whats the difference between applying a " filtered classifier" and "
applying filters first to preprocess ... and... then applying a
The results of classifier accuracy differ a lot ... much improved in
preprocessing first and then applying a classifier algo.
Yes, because the second method involves cheating if the filter is
supervised and makes use of the class labels to construct the filtering
model. For example, a discretization filter would generate
discretization intervals based on both the training and the test data
if you used it in the preprocess panel. It's important that a
supervised filter only has access to the training data when it builds
its internal model.
A particularly drastic example is if attribute selection is performed
in a high-dimensional feature space.