@chkoar WDYT?

I don't have a strong opinion on this. I believe that a robust predictors/ensemble methods module could bring traction to the package. For that reason I would may implement BRF even as a shortcut.

@glemaitre is a PR stalled in scikit-learn where we could contribute to be finished?

it is already merged in #373

I asked because you said .

In that regards this why scikit learn has a PR there.

which one?

scikit-learn/scikit-learn#8732

We wanted also to balanced at each node instead of tree to see the difference.

@glemaitre in each tree? Does this task need modification in the Cython level?

@chkoar, @glemaitre I'm not familiar with imblearn but in sklearn you can balance each tree simply by changing the sample_indices that get passed. This is how I implemented it in scikit-learn/scikit-learn#8732.

So does BalancedBaggingClassifier(max_features="auto") balance at each tree or does it just balance the data once? If the latter, I think it is confusing to call that a "balanced random forest" in the documentation because in the Breiman paper that refers to balancing each tree.

The implementation is a pipeline of a random under sampler with an estimator. So if you pass an estimator which is a tree, it will balance each subset and then fit a tree on each subset. Therefore, BalancedBaggingClassifier become a BalancedRandomForest with max_feature='auto' as mentioned in the documentation.

@glemaitre OK thanks for the clarification