@YuxinB As Adam suggested before, can you add a hyperparameter to turn this function on/off? The kernel method should be the default, but I believe we still need to preserve the standard averaging method.

@YuxinB You can ignore the check errors for the sklearn fork. But can you fix the style and lint errors? Thanks! After that I'll merge the changes.

@YuxinB On second thought, we should not merge the changes until the parameter is implemented. We need to also make sure the forest level predict_proba outputs correctly when testing on an held-out set.

Modify all the default values to what we agreed on.

Merging to facilitate productivity.