A problem about ADASYN

[YazhouZhang0709]

Hi, I read about your ADASYN code and found something confusing. In your ADASYN code, when you create the synthetic samples, you chose a random sample from all the K-neighbors, which including majority samples. However, in the original ADASYN paper, the selected sample from K-neighbors is only chosen from minority class. Could you please check that? Thank you

[glemaitre]

I think you are right. We need to refit the NN and recompute the neighbors before the generation.

[glemaitre]

@chkoar it seems that ADASYN seems to fail even on toy example:

With the bug:

I understand the point that the method will generate more samples where the original samples were difficult to classify. However, it does not make the difference between a noisy sample and a sample in danger. So SMOTE bordeline and SVM seems to be much more powerful in this regard.

The bug introduce was limited this effect since the generated samples were really localized ;)

@chkoar I'll make a PR shortly, if you have any thoughts regarding the analysis, it would be interesting.