Data Balancing Strategies: A Survey of Resampling and Augmentation Methods

TL;DR

The paper addresses the persistent challenge of class imbalance in machine learning by surveying a broad spectrum of data balancing strategies, including traditional oversampling (e.g., SMOTE and its variants), adaptive and undersampling techniques, and modern generative- and ensemble-based approaches. It categorizes methods into synthetic oversampling, adaptive methods, generative models, ensemble and hybrid strategies, undersampling, and near-neighbor based cleaning, detailing their mechanisms, strengths, and typical use cases. A key contribution is the comprehensive mapping of techniques to dataset characteristics (size, feature types, distribution, noise, dimensionality) and the discussion of practical implementations and real-world case studies, culminating in insights about when and how to apply these methods. The paper highlights the growing role of deep generative models (GANs, VAEs) and hybrid methods (e.g., SMOTE with ENN/TL, clustering-based sampling) and points to future directions emphasizing adaptive, data-aware resampling integrated into end-to-end learning pipelines, as well as further exploration of multi-label and clustered data scenarios. Overall, it provides a framework for practitioners to select and combine resampling strategies tailored to specific data properties, aiming to improve classifier performance on imbalanced tasks.

Abstract

Imbalanced data poses a significant obstacle in machine learning, as an unequal distribution of class labels often results in skewed predictions and diminished model accuracy. To mitigate this problem, various resampling strategies have been developed, encompassing both oversampling and undersampling techniques aimed at modifying class proportions. Conventional oversampling approaches like SMOTE enhance the representation of the minority class, whereas undersampling methods focus on trimming down the majority class. Advances in deep learning have facilitated the creation of more complex solutions, such as Generative Adversarial Networks (GANs) and Variational Autoencoders (VAEs), which are capable of producing high-quality synthetic examples. This paper reviews a broad spectrum of data balancing methods, classifying them into categories including synthetic oversampling, adaptive techniques, generative models, ensemble-based strategies, hybrid approaches, undersampling, and neighbor-based methods. Furthermore, it highlights current developments in resampling techniques and discusses practical implementations and case studies that validate their effectiveness. The paper concludes by offering perspectives on potential directions for future exploration in this domain.

Figures (15)

• Figure 1: A tree-structured categorization of data resampling techniques for imbalanced datasets.
• Figure 2: Illustration of the SMOTE algorithm, where synthetic points are generated between minority class instances and their neighbors.
• Figure 3: (a) The initial distribution of the Circle dataset. (b) Minority class instances near the decision boundary (marked as solid squares). (c) Synthetic minority class samples generated near the boundary (depicted as hollow squares).
• Figure 4: Overview of the UnderBagging-like approach. In the confusion information distribution, the symbol size reflects the level of confusion—larger symbols denote instances with greater confusion information iw-smote.
• Figure 5: A visualization of the K-means clustering algorithm IKOTUN2023178.