garve / scikit-bonus Goto Github PK

Hi
This looks great. I see you have lots of very cool functionality implemented. Could you push the SimpleTimeFeatures, please? I assume it would go into skbonus.preprocessing.time?

Thanks!

This is incredible. It works really well with Isotonic regression to add monotonic constraints to the EBM. The one place is doesn't work as expected is with the positivity constraint ymin=0. This restricts the model to a minimum value that is equal to the intercept.

Clearly, this totally makes sense mathematically. What it highlights is that there might be a better way to calculate the intercept so that is not an issue. Possibly, self.mean_ = y.min() in this strictly positive context?

Currently, in the EBM and your mEBM, the simplest solution is to do a constant shift in the y-axis from the minimum value to zero, for each variable, and then update the intercept accordingly:

mins = [np.min(feat) for feat in mebm.outputs_]
mebm.outputs_ = [feat - m for feat, m in zip(mebm.outputs_, mins)]
mebm.mean_ = mebm.mean_ + np.sum(mins)

Anyway, amazing stuff and hopefully these findings/issues/solutions are useful 😄

great code thanks
onoy
punish overestimations by the model 5 times Garve/scikit-bonus: scikit-learn extensions that I find useful.

may you share links to theory behind this code
or just create some blog post pls?