Comprehensive survey maps landscape of techniques for handling imbalanced training datasets

Machine learning practitioners frequently encounter datasets where one class dramatically outweighs others, a structural problem that causes models to favor the majority class and deliver unreliable predictions. A new systematic survey on arXiv catalogs and evaluates the full spectrum of methods available to tackle this imbalance challenge.

The core problem#

Class imbalance remains one of the most stubborn practical issues in machine learning pipelines. When training data contains far more examples of one category than another—say, detecting rare diseases or fraud—standard classifiers tend to overfit toward the prevalent class, resulting in poor generalization and missed minority-class instances. This bias undermines model utility in many real-world applications.

The methods landscape#

The survey organizes balancing techniques into several families:

• Oversampling approaches: Traditional methods like SMOTE (Synthetic Minority Oversampling Technique) and its variants (Borderline SMOTE, K-Means SMOTE, Safe-Level SMOTE) generate synthetic examples to bolster underrepresented classes
• Adaptive and advanced strategies: Newer algorithms such as MWMOTE and AMDO adjust their behavior based on dataset geometry
• Deep generative methods: Generative adversarial networks (GANs), variational autoencoders (VAEs), and diffusion models offer learned approaches to minority-class synthesis
• Undersampling techniques: Methods like NearMiss and Tomek Links reduce majority-class instances to rebalance the dataset
• Hybrid combinations: Approaches merging multiple strategies (SMOTE-ENN, SMOTE-Tomek, SMOTE+OCSVM) aim to leverage complementary strengths
• Ensemble strategies: Integrated approaches such as SMOTEBoost, RUSBoost, Balanced Random Forest, and One-Sided Selection embed balancing into ensemble learning
• Specialized variants: Distinct methods tailored for multi-label scenarios and naturally clustered data

Key insight: No one-size-fits-all solution#

The review's central finding challenges practitioners to abandon hopes for a universal champion. Performance hinges on the specific combination of dataset characteristics (dimensionality, feature types, class overlap, noise levels), the choice of classifier, and how success is measured. No single technique consistently outperforms all others across diverse scenarios—making informed selection crucial.