K-Means SMOTE is an oversampling method for class-imbalanced data. It aids classification by generating minority class samples in safe and crucial areas of the input space. The method avoids the generation of noise and effectively overcomes imbalances between and within classes.

This project is a python implementation of k-means SMOTE. It is compatible with the scikit-learn-contrib project imbalanced-learn.

The implementation is tested under python 3.6 and works with the latest release of the imbalanced-learn framework:

• imbalanced-learn (>=0.4.0, <0.5)
• numpy (numpy>=1.13, <1.16)
• scikit-learn (>=0.19.0, <0.21)

pip install kmeans-smote

Clone this repository and run the setup.py file. Use the following commands to get a copy from GitHub and install all dependencies:

git clone https://github.com/felix-last/kmeans_smote.git
cd kmeans-smote
pip install .

Find the API documentation at https://kmeans_smote.readthedocs.io. As this project follows the imbalanced-learn API, the imbalanced-learn documentation might also prove helpful.

import numpy as np
from imblearn.datasets import fetch_datasets
from kmeans_smote import KMeansSMOTE

datasets = fetch_datasets(filter_data=['oil'])
X, y = datasets['oil']['data'], datasets['oil']['target']

[print('Class {} has {} instances'.format(label, count))
 for label, count in zip(*np.unique(y, return_counts=True))]

kmeans_smote = KMeansSMOTE(
    kmeans_args={
        'n_clusters': 100
    },
    smote_args={
        'k_neighbors': 10
    }
)
X_resampled, y_resampled = kmeans_smote.fit_sample(X, y)

[print('Class {} has {} instances after oversampling'.format(label, count))
 for label, count in zip(*np.unique(y_resampled, return_counts=True))]

Expected Output:

Class -1 has 896 instances
Class 1 has 41 instances
Class -1 has 896 instances after oversampling
Class 1 has 896 instances after oversampling

Take a look at imbalanced-learn pipelines for efficient usage with cross-validation.

K-means SMOTE works in three steps:

1. Cluster the entire input space using k-means [1].
2. Distribute the number of samples to generate across clusters:
   1. Filter out clusters which have a high number of majority class samples.
   2. Assign more synthetic samples to clusters where minority class samples are sparsely distributed.
3. Oversample each filtered cluster using SMOTE [2].

Please feel free to submit an issue if things work differently than expected. Pull requests are also welcome - just make sure that tests are green by running pytest before submitting.

If you use k-means SMOTE in a scientific publication, we would appreciate citations to the following paper:

@article{kmeans_smote,
    title = {Oversampling for Imbalanced Learning Based on K-Means and SMOTE},
    author = {Last, Felix and Douzas, Georgios and Bacao, Fernando},
    year = {2017},
    archivePrefix = "arXiv",
    eprint = "1711.00837",
    primaryClass = "cs.LG"
}

[1] MacQueen, J. “Some Methods for Classification and Analysis of Multivariate Observations.” Proceedings of the Fifth Berkeley Symposium on Mathematical Statistics and Probability, 1967, p. 281-297.

[2] Chawla, Nitesh V., et al. “SMOTE: Synthetic Minority over-Sampling Technique.” Journal of Artificial Intelligence Research, vol. 16, Jan. 2002, p. 321357, doi:10.1613/jair.953.