[BUG]The sampling method of the BRFClassifier is different from the paper

[Chengwen-98]

Describe the bug

Hi.The following is the sampling method of BRF in the paper Using Random Forest to Learn Imbalanced Data：

For each iteration in random forest, draw a bootstrap sample from the minority class. 
Randomly draw the same number of cases, with replacement, from the majority class.

My interpretation is that the minority samples in each sub-training set are selected by bootstrap and each sub-training set is balanced.Then these sub-training sets are given to traditional random forest's trees.
But in the code of imblearn\ensemble\_forest.py\_local_parallel_build_trees ,I found that all minority samples in the training set are used in each sub-training set,the minority samples in the sub-training set of each tree are the same.

[chkoar]

That is correct. I had the impression that we had fixed this. The reference should be removed and the docstring should be updated.
There is an open MR in the scikit-learn repo with an actual implementation from the paper you posted. Although, as you can see, the implemention that we offer though imbalanced-learn is comparable with the original algorithm.

[Chengwen-98]

That is correct. I had the impression that we had fixed this. The reference should be removed and the docstring should be updated.
There is an open MR in the scikit-learn repo with an actual implementation from the paper you posted. Although, as you can see, the implemention that we offer though imbalanced-learn is comparable with the original algorithm.

Thanks for your answer.The reference is still on the website https://pypi.org/project/imbalanced-learn/#id17

[chkoar]

Basically, the reference is everywhere. We should remove the citation from docs and docstrings saying that we have implemented a variation of random forests in order to be adapted in imbalanced data sets.

[Chengwen-98]

Basically, the reference is everywhere. We should remove the citation from docs and docstrings saying that we have implemented a variation of random forests in order to be adapted in imbalanced data sets.

a sudden announcement 😀，thank you for remembering my first issue.

[glemaitre]

a sudden announcement 😀，thank you for remembering my first issue.

Indeed, we can succeed to get the right behaviour only by changing the default. So let's do that.

[glemaitre]

Looking more in detail, I think that I forgot to change the default of bootstrap because it is handled by the sampler.

Note that beforehand, we were passing the minority class entirely to _parallel_bootstrap that later will take a bootstrap as well. So we were indeed taking a bootstrap sample. What change then, is the way that the majority class is handled.