Re: [Wekalist] Adaboost & Threshold curve
by Hans van Rijnberk , Assort Vision, Utrecht
Can anybody tell (irt GUI version of Weka):
Why is the number instances for the threshold curve reduced when I use
Adaboost with logistic regression. This is even so when an hold out set is
used for testing, so it can't be because the number of independent estimates
I myself use the ROC analysis (TP vs FP with Area Under the ROC). When I use
logistic regression without Adaboost I get [ROC instances] = [original
number of instances - 2].
This is strange, because (TP,FP)= 0,0 and 1,1 are explicite points of a ROC.
The lowest threshold value should be equivalent to the lowest probability of
the positive class and represents (TP,FP) = (0,0). Also, the highets
threshold value should be equivalent to the highest probability of the
positive class and represents (TP,FP) = (1,1).
Hans van Rijnberk
(machine vision software & information services)
3524 KM Utrecht, the Netherlands
tel 031 30 2889531