[MRG] Add the SMOTE-RSB oversampling technique

[zoj613]

Reference Issue

Related to checklist in #105

What does this implement/fix? Explain your changes.

Adds the SMOTE-RSB oversampling technique.

Any other comments?

NA

[pep8speaks]

Hello @zoj613! Thanks for updating this PR. We checked the lines you've touched for PEP 8 issues, and found:

• In the file imblearn/over_sampling/_smote.py:

Line 609:17: W503 line break before binary operator
Line 857:89: E501 line too long (91 > 88 characters)
Line 1467:46: W504 line break after binary operator

Comment last updated at 2021-02-05 00:10:14 UTC

[zoj613]

Not sure why there so many build failures. The tests pass locally.

[zoj613]

Ping @glemaitre @hayesall @chkoar . I think this is good for a review.

[zoj613]

ping @glemaitre

[glemaitre]

@zoj613 I think that we should prioritize which SMOTE variants to include in imbalanced-learn. I think that the benchmark done there https://github.com/analyticalmindsltd/smote_variants/ is quite interesting.

I would be more in favor on implementing the 6 versions mentioned in the comment: analyticalmindsltd/smote_variants#14 (comment)

I need to update the issue in this regard.

[glemaitre]

I look quickly at the paper. It seems that the similarity is used to find some neighbours.
I am not entirely sure but a NearestNeighbors with a given radius (i.e. the similarity-value (distance normalized by the max distance) would be a way to get a faster implementation. Managing categorical and numerical value would be to implement different distance (e.g. preocmputed ValueDifferenceMetric as in SMOTEN.