"SKEWED gamma like Non-negative pdf modeling= poor learner performance"

frito

Hi

I have noticed that if a label attribute has a  non negative highly skewed distribution the performance of learners such as Neural net is very poor compared to symmetrical (about zero) pdf problems.

Any way to tweak/bias a learner for such a distribution?

thx

IngoRM

Hi,

maybe it helps if you normalize the label before training and de-normalize the prediction after model application... I am not sure if this helps (not using NN a lot myself...)

Cheers,
Ingo