Claudio Lucchesi wrote:
Why do I get /slight/ differences in classified events
trees and confusion matrices, when doing classifications with J48?
Probably because of "fractional" instances that are due to J48's
treatment of missing values. For example, 0.8 percent of a training instance
may end up in a "negative" leaf, and 0.2 percent in a "positive" leaf.
classification assigned to this instance will be "negative". However,
in the decision tree output itself, you will find 0.8 added to the counts
for the corresponding positive leaf and 0.2 added to the counts for the
corresponding negative leaf.
Besides, I suggest that sensitivity and specificity be
outputted by Weka.
The are included in the default output shown by the explorer and called
"precision" and "recall" respectively. From the command line you can
get this extra output using the -i flag.