feature request: BoosterPipeline about scikit-lego

[koaning]

This is a good question. I might argue that:

1. this could be a setting
2. the base behaviour should probably be that you don't use the predictions from a previous model, rather that you the same features and that you had before (depending on other transformations)

[MaxHalford]

Not sure this is what you're looking for but I wrote a generic gradient library called starboost from which you might want to inspire yourself!

[koaning]

The idea here is to implement a pipeline like;

from sklego.transformers import RandomAdder

from sklearn.preprocessing import StandardScaler
from sklearn.linear_model import LogisticRegression
from sklearn.ensemble import RandomForest 

...

mod = BoosterPipeline([
    ("scale", StandardScaler()),
    ("random_noise", RandomAdder()),
    ("global_model", LogisticRegression(solver='lbfgs')), 
    ("residual_model", RandomForest())
])

...

It's more of a UI thing. The global_model will first run and then the BoosterPipeline should recognise that the residuals need to be picked up.

[MaxHalford]

So it is still like some form of stacking? Are you going to reuse the features that the global used, or are you going to base yourself on the predictions?

[koaning]

I am wondering if this feature might break up the workflow too much. Instead it might also work if we just have a Booster-model that accepts many different models.

[MBrouns]

I tried to build something like this quite a while back and remember it took a lot of changes in the normal pipeline behaviour. A Booster meta-model might be simpler