A Comprehensive Review of Data Balancing Techniques in Machine Learning: Methods, Challenges, and Future Directions

The problem of class imbalance is an important problem in machine learning (ML), leading to bias in the trained models and poor accuracy, especially for the minority classes. This review performs a thorough characterization of Under-sampling, over-sampling, algorithm-level, and their hybrid techniques. It evaluates popular techniques such as Random Under-sampling (RUS), Synthetic Minority Over-sampling Technique (SMOTE), and Cost-Sensitive Learning (CSL), and explores advanced hybrid strategies like SMOTE-ENN and SMOTE+Tomek Links. The effectiveness of these approaches is compared across a wide range of datasets and ML techniques, highlighting their advantages, trade-offs, and practical limitations, including risks of overfitting, information loss, and computational complexity. This paper also discusses challenges in multi-class imbalance scenarios and high-dimensional data. Finally, it discusses future research directions including balancing with Deep Learning (DL), AutoML, and fairness-aware modeling. It provides a reliable overview of the state of the art by presenting a summary of literature and details gaps, serving as a blueprint for researchers and potential users who are interested in accurate strategies to solve the class imbalance problem in various real-world applications.