Class and sample weights are not taken into account when calibrating probabilities about libsvm HOT 5 OPEN

from libsvm.

Yes. Weights have been considered in gettting dec values that are inputs of
sigmoid_train

Peter Prettenhofer writes:

It seems to me that sigmoid_train doesnt take weights into account.

https://github.com/cjlin1/libsvm/blob/master/svm.cpp#L1705

—
Reply to this email directly or view it on GitHub.*

from libsvm.

but dont we need to take weight also into account when doing Platt scaling? Platt scaling is just a logistic regression with one input (the confidence scores of the SVM), right? Training this model should be done by weighting log loss based on the sample weights.

from libsvm.

Sorry for the late reply. We are studying this issue, but it will take some
time.
Peter Prettenhofer writes:

but dont we need to take weight also into account when doing Platt scaling?
Platt scaling is just a logistic regression with one input (the confidence
scores of the SVM), right? Training this model should be done by weighting log
loss based on the sample weights.

—
Reply to this email directly or view it on GitHub.*

from libsvm.

Hmm ... are there other functions that actually use sample weights? I don't even see them described in the README.

from libsvm.