You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
For some tasks, with a lot of features, it might be handy to let Timbl select the top N features, based on the current weight, and use only those to build the tree. Also an implicit -mI for all the other features.
e.g a --cutoff 1000, would select the 1000 'best ranked' features. Assuming more than 1000 are available :)
comment welcome....
The text was updated successfully, but these errors were encountered:
small note: The Feature weights are at first calculated using ALL available features.
After choosing the N best, the weights have to be recalculated for only those N.
Otherwise surprises might happen.
for instance: when restoring an IB1 tree from a file, Timbl calculates the weights based on the info in the file. (Also NOT taking into account the ignored features.)
These weights should match those of an tree prior to storing.
So the scheme is:
calculate the weights
select the N best
recalculate the weights, with a -mI:k for all all NOT in the N best.
For some tasks, with a lot of features, it might be handy to let Timbl select the top N features, based on the current weight, and use only those to build the tree. Also an implicit -mI for all the other features.
e.g a --cutoff 1000, would select the 1000 'best ranked' features. Assuming more than 1000 are available :)
comment welcome....
The text was updated successfully, but these errors were encountered: