feature request: ensemble model that takes p(class | model) into account about scikit-lego

[koaning]

700 downloads cant be wrong.

https://link.springer.com/chapter/10.1007/978-3-319-00651-2_19

[koaning]

I'll make it slightly clearer what I am thinking about here.

Consider that we have an ensemble of three models. They each have a confusion matrix and at the time of inference they each come with a prediction for class B.

Do we really trust what model A has to say here? When it predicts class B then it is often wrong.

So maybe ... we need a different way of reweighting here. An idea I've had for a while is to try to reweigh the weights using the confusion matrix. There is even something bayesian about it.

So what I am proposing here is similar to soft voting in the VotingClassifier but slightly different. I would like the algorithm to recognize that is should not listen demographically to all the models if it turns out that certain models are unreliable for certain decisions.