We have proposed a new oversampling algorithm called GB-SMOTE, which can circumvent the deficiency of WK-SMOTE [1] (and SMOTE as well as its variants) caused by randomly selecting some minority class samples. To design this oversampling method, we first design a simple grouping scheme that can divide the minority class into different groups by using slack variables. This grouping scheme has established a theoretical basis for selecting valuable minority class samples. Moreover, it also provides a new explanation for the poor performance of SVM on imbalanced data sets. Second, a reasonable samples selection scheme has been designed, which can avoid generating new samples in the overlapping region, and an effective samples generation scheme is proposed to generate high-quality new samples. Subsequently, an effective oversampling method GB-SMOTE is proposed. The idea of GB-SMOTE (partially selecting valuable samples) can also be applied to preprocess biased learners or to modify the input distribution in cases of a limited amount of data in semi-supervised learning. The experimental results indicate the effectiveness of the sample selection scheme and sample generation scheme in GB-SMOTE. Besides, the experiment on the real-world datasets shows that compared with all of the benchmark algorithms in both homologous and heterologous groups, GB-SMOTE outperforms them, especially on the data sets with a high imbalance ratio.

If you find this repository helpful in your work or research, we would greatly appreciate citations to the following paper:

@article{REN20220822,
title = {Grouping-based Oversampling in Kernel Space for Imbalanced Data Classification},
journal = {Pattern Recognition},
volume = {133},
pages = {108992},
year = {2023},
issn = {0031-3203},
doi = {https://doi.org/10.1016/j.patcog.2022.108992},
url = {https://www.sciencedirect.com/science/article/pii/S0031320322004721},
author = {Jinjun Ren and Yuping Wang and Yiu-ming Cheung and Xiao-Zhi Gao and Xiaofang Guo}
}

Our GB-SMOTE implementation requires following dependencies:

• python (>=3.7)
• numpy (>=1.11)
• scipy (>=0.17)
• scikit-learn (>=0.21)

git clone https://github.com/JinJunRen/GB-SMOTE

GB-SMOTE.py

demorun.py

In this python script we provided an example of how to use our implementation of GB-SMOTE methods to perform classification.

python demorun.py -data ./dataset/moon_1000_100_2.csv -n 5 
or
python demorun.py -data ./dataset/moon_1000_200_4.csv -n 5

##Dataset links: Knowledge Extraction Based on Evolutionary Learning (KEEL).

• [1] Mathew J, Pang C K, Luo M, et al. Classification of imbalanced data by oversampling in kernel space of support vector machines[J]. IEEE transactions on neural networks and learning systems, 2017, 29(9): 4065-4076.