Question: Adding offset variable during training

[HoShilo]

While training an EBM regressor/classifier, is it possible to integrate an offset variable as in GLM/GAM? By offset, I mean a fixed effect variable that is added to the additive formula. For instance, in XGBOOST, this is implemented with the base_margin argument.

I looked into the init_score of the fit method, but I don't think it serves the same purpose. From my understanding, it doesn't pass through the link function. It is easy to bypass this with identity or log link functions (just subtract or divide y before the fit by offset or exp(offset) respectively), but with classifiers (logit link), it is less straightforward how to approach this.

[paulbkoch]

Hi @HoShilo --

init_score is the right approach, and it gets added to the additive scores from the other features before applying the inverse link during prediction. If you had a more complicated model, then you would need to use the init_score on the predict function during prediction in addition to using it during fitting. For adding a constant base_margin though, I would just add that value to the intercept_ attribute on the EBM after fitting, which will have the same effect, but is simpler.