README.md

Do we need rebalancing strategies? A theoretical and empirical study around SMOTE and its variants.

Repository for Do we need rebalancing strategies? A theoretical and empirical study around SMOTE and its variants paper.

In praticular, you will find code to reproduce the paper experiments as well as an nice implementation of our new and efficient strategy for your projects.

⭐ Table of Contents

• Getting Started
• Data sets
• Acknowledgements

⭐ Getting Started

If you want to reproduce our paper experiments:

• the notebooks here and here reproduce the experiments
• thise code contains implementation the protocols used for the numerical experiments of our article.

In order to use our MGS strategy:

• this notebook illustrates how to use it
• the strategy is implemented here

⭐ Data sets

The data sets of used for our article should be dowloaded inside the data/externals folder. The data sets are available at the followings adresses :

• Pima
• Phoneme : https://github.com/jbrownlee/Datasets/blob/master/phoneme.csv
• Abalone : https://archive.ics.uci.edu/dataset/1/abalone
• Wine : https://archive.ics.uci.edu/dataset/186/wine+quality
• Haberman : https://archive.ics.uci.edu/dataset/43/haberman+s+survival
• Yeast : https://archive.ics.uci.edu/dataset/110/yeast
• Vehicle : https://archive.ics.uci.edu/dataset/149/statlog+vehicle+silhouettes
• Ionosphere : https://archive.ics.uci.edu/dataset/52/ionosphere
• Breast cancer Wisconsin : https://archive.ics.uci.edu/dataset/15/breast+cancer+wisconsin+original
• CreditCard : https://www.kaggle.com/datasets/mlg-ulb/creditcardfraud
• MagicTel : https://www.openml.org/d/44125
• California : https://www.openml.org/d/44090
• House_16H : https://openml.org/d/821

Table 2 from the paper :

Strategy	None	CW	RUS	ROS	NM1	BS1	BS2	SMOTE	CV SMOTE	MGS (d+1)
CreditCard (0.2%)	$0.966$	$0.967$	0.970	$0.935$	$0.892$	$0.949$	$0.944$	$0.947$	$0.954$	$0.952$
Abalone (1%)	$0.764$	$0.748$	$0.735$	$0.722$	$0.656$	$0.744$	$0.753$	$0.741$	$0.791$	0.802
Phoneme (1%)	$0.897$	$0.868$	$0.868$	$0.858$	$0.698$	$0.867$	$0.869$	$0.888$	0.924	$0.915$
Yeast (1%)	$0.925$	$0.920$	$0.938$	$0.908$	$0.716$	$0.949$	$0.954$	0.955	$0.942$	$0.945$
Wine (4%)	$0.928$	$0.925$	$0.915$	$0.924$	$0.682$	$0.933$	$0.927$	$0.934$	$0.938$	0.941
Pima (20%)	$0.798$	0.808	$0.799$	$0.790$	$0.777$	$0.793$	$0.788$	$0.789$	$0.787$	$0.787$
Haberman (10%)	$0.708$	$0.709$	$0.720$	$0.704$	$0.697$	$0.723$	$0.721$	$0.719$	$0.742$	0.744
MagicTel (20%)	$0.917$	$0.921$	$0.917$	0.922	$0.649$	$0.920$	$0.905$	$0.921$	$0.919$	$0.913$
California (1%)	$0.887$	$0.877$	$0.880$	$0.883$	$0.630$	$0.885$	$0.874$	$0.906$	$0.916$	0.923

⭐ Acknowledgements

This work was done through a partenership between Artefact Research Center and the Laboratoire de Probabilités Statistiques et Modélisation (LPSM) of Sorbonne University.

If you find the code usefull, please consider citing us :

@article{sakho2024theoretical,
  title={Theoretical and experimental study of SMOTE: limitations and comparisons of rebalancing strategies},
  author={Sakho, Abdoulaye and Scornet, Erwan and Malherbe, Emmanuel},
  journal={arXiv preprint arXiv:2402.03819},
  year={2024}
}