Imbalanced-learn: A Python Toolbox to Tackle the Curse of Imbalanced Datasets in Machine Learning

Guillaume LemaitreFernando NogueiraChristos K. Aridas

article2017JMLR2,621 citations

Presents imbalanced-learn, a Python library integrated with scikit-learn that provides standardized implementations of over-sampling, under-sampling, and ensemble algorithms to effectively train machine learning models on skewed class distributions.

Listen

Real-world datasets in fields such as fraud detection, medical diagnosis, and bioinformatics often contain far fewer examples of one class than others. This imbalance disrupts standard machine learning algorithms that assume roughly equal class sizes or costs, leading to poor performance on the minority class. The problem ranks among the top ten challenges in data mining and pattern recognition.

The article set out to introduce and describe an open-source Python toolbox that supplies a range of established methods for correcting class imbalance. The toolbox aims to fill a gap in the Python ecosystem by offering tools that integrate directly with widely used libraries.

The authors built the imbalanced-learn package around four categories of techniques: under-sampling, over-sampling, combined sampling, and ensemble approaches. Development followed scikit-learn conventions, relied solely on numpy, scipy, and scikit-learn, and incorporated unit tests achieving 99 percent coverage, continuous integration, and community review processes. At the time of writing, the repository attracted roughly 2,000 visits and 300 unique visitors each week.

The toolbox supplies fixed and cleaning under-sampling routines, random and SMOTE-based over-sampling with several variants, combinations that pair SMOTE with cleaning steps, and two ensemble methods that reuse majority-class samples across multiple balanced subsets. All samplers expose a consistent fit, sample, and fit-sample interface, and a Pipeline class allows seamless chaining with scikit-learn transformers and estimators.

These capabilities give practitioners a single, maintained Python resource for balancing data before model training, reducing the need to switch languages or implement methods from scratch. Integration with scikit-learn lowers adoption barriers and supports reproducible workflows in production environments.

The authors plan to add further prototype-selection and generation methods along with additional user guides. Users should monitor the GitHub repository for these extensions and test the current release on their own datasets before large-scale deployment.

The description focuses on implementation and design rather than new empirical benchmarks, so performance gains will vary by data set and downstream classifier. Readers should verify results on representative samples and consider the MIT license and scikit-learn-contrib status when evaluating long-term maintenance.

ShoppeAI Universitat de Girona Université Bourgogne Franche-Comté University of Patras ensemble learning pattern recognition python class imbalance imbalanced datasets ensemble learning

arXiv: 1609.06570 scikit-learn-contrib/imbalanced-learn

Paper: SMOTE: Synthetic Minority Over-sampling Technique, N. Chawla et al. (2002). This seminal paper introduces SMOTE, the foundational synthetic oversampling algorithm implemented and extended throughout the imbalanced-learn toolbox.
Paper: Scikit-learn: Machine Learning in Python, Fabian Pedregosa et al. (2011). It outlines the core design, API conventions, and pipelining standards that imbalanced-learn strictly adopts to ensure seamless compatibility within the Python ecosystem.
Paper: Addressing the Curse of Imbalanced Training Sets: One-Sided Selection, M. Kubát et al. (1997). This foundational work introduces One-Sided Selection and Tomek-link cleaning concepts that serve as primary under-sampling routines implemented in the toolbox.
Paper: The foundations of cost-sensitive learning, Charles Elkan (2001). This work establishes the theoretical foundation linking class rebalancing to cost-sensitive decision-making and optimal classification thresholds.
Paper: The relationship between Precision-Recall and ROC curves, Jesse Davis et al. (2006). It provides essential evaluation theory demonstrating why standard ROC analysis can be misleading under heavy class skew and why precision-recall dynamics must be considered.

Paper: SMOTE for Learning from Imbalanced Data: Progress and Challenges, Marking the 15-year Anniversary, Alberto Fernández et al. (2018). This comprehensive 15-year retrospective systematically evaluates dozens of SMOTE variants and modern challenges, extending the survey of techniques implemented in the imbalanced-learn package.
Paper: Class-Balanced Loss Based on Effective Number of Samples, Yin Cui et al. (2019). This work advances beyond traditional heuristic resampling techniques by formulating a loss-reweighting framework based on the theoretical effective number of samples.
Paper: Learning Imbalanced Datasets with Label-Distribution-Aware Margin Loss, Kaidi Cao et al. (2019). It extends imbalanced learning to deep neural networks through margin loss adjustments and deferred re-balancing schedules that address the shortcomings of classical resampling.
Paper: A systematic study of the class imbalance problem in convolutional neural networks, Mateusz Buda et al. (2017). This empirical study evaluates how classical resampling remedies—like those packaged in imbalanced-learn—perform when applied directly to modern convolutional neural network architectures.

Cover for Imbalanced-learn: A Python Toolbox to Tackle the Curse of Imbalanced Datasets in Machine Learning

Abstract

Imbalanced-learn is an open-source python toolbox aiming at providing a wide range of methods to cope with the problem of imbalanced dataset frequently encountered in machine learning and pattern recognition. The implemented state-of-the-art methods can be categorized into 4 groups: (i) under-sampling, (ii) over-sampling, (iii) combination of over- and under-sampling, and (iv) ensemble learning methods. The proposed toolbox only depends on numpy, scipy, and scikit-learn and is distributed under MIT license. Furthermore, it is fully compatible with scikit-learn and is part of the scikit-learn-contrib supported project. Documentation, unit tests as well as integration tests are provided to ease usage and contribution. The toolbox is publicly available in GitHub: this https URL.

Topics

ensemble learning pattern recognition python matlab

Keywords

class imbalance imbalanced datasets ensemble learning Tomek links nearest neighbor class distribution Python toolbox

1 Introduction
2 Project management
3 Implementation design
4 Implemented methods
4.1 Notation and background
4.2 Under-sampling
4.3 Over-sampling
4.4 Combination of over- and under-sampling
4.5 Ensemble learning
5 Future plans and conclusion
References

Knowls

Knowl 1 — API Architecture and Sampler Interface of imbalanced-learn
model/method
The imbalanced-learn Python library provides a modular interface for class imbalance mitigation designed to integrate with the scikit-learn ecosystem. Resampling algorithms derive from dedicated sampler classes that expose three primary methods:
- fit(X, y): Analyzes the training feature matrix $X$ and target label vector $y$ to compute internal parameters and decision statistics necessary for resampling.
- sample(X, y): Applies the computed resampling strategy to $X$ and $y$ , returning balanced feature and target arrays according to a target balancing ratio.
- fit_sample(X, y): A convenience method that executes fit(X, y) followed immediately by sample(X, y) on the input dataset.
To allow integration with standard learning workflows, the library provides a custom Pipeline class subclassed from scikit-learn that allows samplers to be chained sequentially with feature transformers and estimators.
Knowl 2 — Balancing Ratio for Imbalanced Datasets
definition
For a binary classification dataset $\chi$ partitioned into a minority class subset $\chi_{\text{min}}$ and a majority class subset $\chi_{\text{maj}}$ , the dataset balancing ratio $r_\chi$ is defined as:

$r_\chi = \frac{|\chi_{\text{min}}|}{|\chi_{\text{maj}}|}$

where $|\cdot|$ denotes set cardinality. A data balancing procedure transforms $\chi$ into a resampled dataset $\chi_{\text{res}}$ with a targeted balancing ratio $r_{\chi_{\text{res}}}$ by adjusting the relative sizes of the minority and majority classes.
Knowl 3 — Fixed vs. Cleaning Under-Sampling Taxonomy
model/method
Under-sampling reduces the majority class sample size $|\chi_{\text{maj}}|$ to balance class distributions. Implemented under-sampling approaches are categorized into two paradigms:
- Fixed under-sampling: Directly downsamples the majority class to satisfy a specified balancing ratio $r_{\chi_{\text{res}}}$ . Selection criteria include uniform random sampling, cluster centroid substitution, nearest neighbour rules (e.g., NearMiss variants), and classification accuracy ranking via instance hardness thresholds.
- Cleaning under-sampling: Does not target a predetermined balancing ratio $r_{\chi_{\text{res}}}$ , but instead refines the feature space by removing noisy, ambiguous, or borderline majority samples based on local neighborhood criteria. Methods include Condensed Nearest Neighbours (CNN), Edited Nearest Neighbours (ENN), One-Sided Selection (OSS), Neighbourhood Cleaning Rule (NCR), and Tomek links.
Knowl 4 — Over-Sampling Techniques for Minority Class Augmentation
model/method
Over-sampling increases the number of minority samples in $\chi_{\text{min}}$ until a desired balancing ratio $r_{\chi_{\text{res}}}$ is reached. The toolbox implements two main classes of over-sampling:
- Random over-sampling: Randomly duplicates existing instances from $\chi_{\text{min}}$ with replacement until the target class balance is achieved.
- SMOTE (Synthetic Minority Over-sampling Technique): Synthesizes new feature vectors by interpolating along the line segments joining minority class instances and their nearest neighbours within $\chi_{\text{min}}$ .
- SMOTE variants: Specialized heuristics including Borderline-SMOTE (Borderline 1 and Borderline 2, which selectively synthesize points near decision boundaries) and SVM-SMOTE (which generates synthetic instances using support vectors established by a Support Vector Machine).
Knowl 5 — Hybrid Sampling Combining Over-Sampling and Cleaning Under-Sampling
model/method
Synthetic over-sampling via SMOTE can induce overfitting and produce noisy samples within majority class regions. Hybrid methods mitigate this by applying a cleaning under-sampling stage immediately following SMOTE generation:
- SMOTE + Tomek links: First expands the minority class using SMOTE, then detects and removes all Tomek links (pairs of nearest neighbor instances belonging to opposing classes) to clarify decision boundaries.
- SMOTE + Edited Nearest Neighbours (ENN): First expands the minority class using SMOTE, then applies ENN to eliminate any instance whose class label disagrees with the majority class of its nearest neighbours.
Knowl 6 — Ensemble Under-Sampling via EasyEnsemble and BalanceCascade
model/method
To prevent information loss inherent to discarding majority instances during under-sampling, ensemble under-sampling creates multiple balanced subsets to train a classifier collection:
- EasyEnsemble: Repeatedly draws independent random under-sampled subsets of the majority class $\chi_{\text{maj}}$ , pairs each with the entire minority class $\chi_{\text{min}}$ to construct multiple balanced subsets with ratio $r_{\chi_{\text{res}}}$ , and trains an ensemble of learners.
- BalanceCascade: Constructs balanced training subsets sequentially using feedback from intermediate classifiers. At each stage, a classifier is trained on a balanced subset; majority class samples correctly classified by the current ensemble are removed from future consideration, while misclassified majority samples are retained for subsequent sampling stages.

Coverage note — No substantial contributed material was omitted. The software's project infrastructure details (e.g., Travis CI, test coverage percentage, and documentation tools) were omitted as standard non-methodological engineering metadata.

References

1.G. E. Batista, A. L. Bazzan, and M. C. Monard. Balancing training data for automated annotation of keywords: a case study. In WOB, pages 10–18, 2003.
2.N. V. Chawla, K. W. Bowyer, L. O. Hall, and W. P. Kegelmeyer. SMOTE: synthetic minority over-sampling technique. Journal of artificial intelligence research, pages 321–357, 2002.
3.A. Dal Pozzolo, O. Caelen, S. Waterschoot, and G. Bontempi. Racing for unbalanced methods selection. In International Conference on Intelligent Data Engineering and Automated Learning, pages 24–31. Springer, 2013.
4.H. Han, W.-Y. Wang, and B.-H. Mao. Borderline-smote: a new over-sampling method in imbalanced data sets learning. In International Conference on Intelligent Computing, pages 878–887. Springer, 2005.
5.P. Hart. The condensed nearest neighbor rule. Information Theory, IEEE Transactions on, 14(3):515–516, May 1968.
6.H. He and E. Garcia. Learning from imbalanced data. Knowledge and Data Engineering, IEEE Transactions on, 21(9):1263–1284, 2009.
7.M. Kubat, S. Matwin, et al. Addressing the curse of imbalanced training sets: one-sided selection. In International Conference in Machine Learning, volume 97, pages 179–186. Nashville, USA, 1997.
8.M. Kuhn. Caret: classification and regression training. Astrophysics Source Code Library, 1:05003, 2015.
9.J. Laurikkala. Improving identification of difficult small classes by balancing class distribution. Springer, 2001.
10.X.-Y. Liu, J. Wu, and Z.-H. Zhou. Exploratory undersampling for class-imbalance learning. IEEE Transactions on Systems, Man, and Cybernetics, Part B (Cybernetics), 39(2):539–550, 2009.
11.I. Mani and I. Zhang. knn approach to unbalanced data distributions: a case study involving information extraction. In Proceedings of Workshop on Learning from Imbalanced Datasets, 2003.
12.H. M. Nguyen, E. W. Cooper, and K. Kamei. Borderline over-sampling for imbalanced data classification. International Journal of Knowledge Engineering and Soft Data Paradigms, 3(1):4–21, 2011.
13.F. Pedregosa, G. Varoquaux, A. Gramfort, V. Michel, B. Thirion, O. Grisel, M. Blondel, P. Prettenhofer, R. Weiss, V. Dubourg, et al. Scikit-learn: Machine learning in python. Journal of Machine Learning Research, 12(Oct):2825–2830, 2011.
14.R. C. Prati, G. E. Batista, and M. C. Monard. Data mining with imbalanced class distributions: concepts and methods. In Indian International Conference Artificial Intelligence, pages 359–376, 2009.
15.M. Rastgoo, G. Lemaitre, J. Massich, O. Morel, F. Marzani, R. Garcia, and F. Meriaudeau. Tackling the problem of data imbalancing for melanoma classification. In Bioimaging, 2016.
16.M. R. Smith, T. Martinez, and C. Giraud-Carrier. An instance level analysis of data complexity. Machine learning, 95(2):225–256, 2014.
17.S. C. Sonnenburg, S. Henschel, C. Widmer, J. Behr, A. Zien, F. de Bona, A. Binder, C. Gehl, V. Franc, et al. The SHOGUN machine learning toolbox. Journal of Machine Learning Research, 11(Jun):1799–1802, 2010.
18.I. Tomek. Two modifications of CNN. Systems, Man, and Cybernetics, IEEE Transactions on, 6:769–772, 1976.
19.L. Torgo. Data mining with R: learning with case studies. Chapman & Hall/CRC, 2010.
20.D. L. Wilson. Asymptotic properties of nearest neighbor rules using edited data. Systems, Man and Cybernetics, IEEE Transactions on, (3):408–421, 1972.
21.Q. Yang and X. Wu. 10 challenging problems in data mining research. International Journal of Information Technology & Decision Making, 5(04):597–604, 2006.

Citation

MLA

Lemaître, G., et al. “Imbalanced-learn: A Python Toolbox to Tackle the Curse of Imbalanced Datasets in Machine Learning”. Journal of Machine Learning Research, vol. 18, no. 17, 2017, pp. 1–5, https://www.jmlr.org/papers/v18/16-365.html.

APA

Lemaître, G., Nogueira, F., & Aridas, C. K. (2017). Imbalanced-learn: A Python Toolbox to Tackle the Curse of Imbalanced Datasets in Machine Learning. Journal of Machine Learning Research, 18(17), 1–5. https://www.jmlr.org/papers/v18/16-365.html

Chicago

Lemaître, G., F. Nogueira, and C. K. Aridas. 2017. “Imbalanced-learn: A Python Toolbox to Tackle the Curse of Imbalanced Datasets in Machine Learning”. Journal of Machine Learning Research 18 (17): 1–5. https://www.jmlr.org/papers/v18/16-365.html.

Harvard

Lemaître, G., Nogueira, F. and Aridas, C.K. (2017) “Imbalanced-learn: A Python Toolbox to Tackle the Curse of Imbalanced Datasets in Machine Learning”, Journal of Machine Learning Research, 18(17), pp. 1–5. Available at: https://www.jmlr.org/papers/v18/16-365.html.

Vancouver

1. Lemaître G, Nogueira F, Aridas CK (2017) Imbalanced-learn: A Python Toolbox to Tackle the Curse of Imbalanced Datasets in Machine Learning. Journal of Machine Learning Research 18:1–5

BibTeX

@article{JMLR:v18:16-365,
  author  = {Guillaume  Lema{{\^i}}tre and Fernando Nogueira and Christos K. Aridas},
  title   = {Imbalanced-learn: A Python Toolbox to Tackle the Curse of Imbalanced Datasets in Machine Learning},
  journal = {Journal of Machine Learning Research},
  year    = {2017},
  volume  = {18},
  number  = {17},
  pages   = {1--5},
  url     = {http://jmlr.org/papers/v18/16-365.html}
}

Metadata:DOI registry

Source Code

This paper has an official code repository available. Click below to access the source code.

View Repository

Access the Paper

This paper is available from its original source. Click below to access the PDF.

Open PDF

License: https://creativecommons.org/licenses/by/4.0/

Imbalanced-learn: A Python Toolbox to Tackle the Curse of Imbalanced Datasets in Machine Learning

Abstract

Topics

Keywords

Table of Contents

Knowls

Knowl 1 — API Architecture and Sampler Interface of imbalanced-learn

Knowl 2 — Balancing Ratio for Imbalanced Datasets

Knowl 3 — Fixed vs. Cleaning Under-Sampling Taxonomy

Knowl 4 — Over-Sampling Techniques for Minority Class Augmentation

Knowl 5 — Hybrid Sampling Combining Over-Sampling and Cleaning Under-Sampling

Knowl 6 — Ensemble Under-Sampling via EasyEnsemble and BalanceCascade

References

Citation

BibTeX

Source Code

Access the Paper