Imbalanced Regression: Definition, Impacts and Solutions

Data quality and availability are central challenges in statistical and machine learning, particularly when modeling rare events. This work investigates the use of synthetic data generation as a means to enhance the performance of standard learning approaches in imbalanced settings, covering both classification and regression tasks. Generating realistic synthetic tabular data is inherently difficult, as it requires faithfully capturing complex inter-variable relationships. While generation can be performed directly in the original feature space, we demonstrate that representation learning offers significant advantages by uncovering non-linear correlations and improving the overall quality of the generated samples. We first provide a comprehensive analysis of data imbalance, characterizing its impact both empirically and theoretically. We then propose several data generation strategies designed to mitigate this imbalance and improve model accuracy and generalization. By extending imbalanced data management beyond binary classification to regression and more complex real-world settings where rare outcomes are often the most consequential, this work opens new perspectives for statistical modeling and machine learning in high-stakes applications.