What is the implemented strategy for multiclass resampling?

[SantiagoEG]

Hello,

firstly I would like to apprecite the contributors´ for providing this great library.
I am using imbalance-learn for my Phd and I have a crucial question to clarify the methodology inherently used by the library.

From the literature, I have found two main strategies to apply resampling algorithms, designed for binary problems, to multiclass problems: OnevsOne and OnvsAll.

I need to know what multiclass strategy has been implemented.
Thanks you.

[glemaitre]

The multiclass strategy implemented are the one which is given in the original paper.
I would redirect you to the literature for each methods.

However, this is true that we should probably make a note to avoid going back to the literature.

[glemaitre]

Additionally, we need an entry in the User Guide.

[JACKYLUO1991]

Hello, I just started using this great toolbox, really did not see how to carry out multiple types of resampling, I would like to know how to use，thanks very much！

[glemaitre]

@actionLUO not related to the current issue. Please open a new issue following the contributing guide to raise issue.

[SantiagoEG]

Hello,

@glemaitre, sorry for the inconvinences, I think I didn´t express properly my question.

In the original literature of SMOTE (https://www.jair.org/media/953/live-953-2037-jair.pdf) the authors tested SMOTE algorithm over several binary problems (see Section 5.1 Datasets). The resampling algorithms is normally presented for two-classes problems.

I am using these algorithms for a multiclass problem. Briefly, two descomposition techniques for oversampling techniques are applied in multiclass problems to simplify them in several binary problems:

• OneVsOne: Two classes are selected and the resampling algorithm is applied to the samples belonging to these two classes. Then, the new samples generated for each pair of classes are added to the original dataset.

• OneVsRest: the multiclass problem is simplified labeling all samples belonging to one classes as 1, and the rest as 0. Then the resampling algoirthm is applied in this way for each class and the new samples are added to the original dataset.

(more literature about this: http://ieeexplore.ieee.org/document/6170916/ and http://www.sciencedirect.com/science/article/pii/S0950705113000300)

My doubt is about this. When I use SMOTE for a multiclass problem I use the dictionary. When I do this, what descomposition techniques is employed?

Thank you for your contributions.

[glemaitre]

one-vs-rest as you can extrapolate from the original paper: "Nevertheless, the SMOTE technique can be applied to a multiple class problem as well by specifying what class to SMOTE for."

[SantiagoEG]

Thank you very much. Really, the authors do not explained the multiclass adaptation clearly.
¡Issue Solved!

[glemaitre]

@SantiagoEG I agree, this is an extrapolation :)