NO-IGS: Nonlinear Oversampling Based on Informed Granular-ball Space for Imbalanced Regression
Hao Zhou, Xiaofeng Liao, Qingguo Lü, Ponnuthurai Nagaratnam Suganthan · IEEE Transactions on Artificial Intelligence · 2026
In this paper, we consider the problem of imbalanced regression. This problem is prevalent in real-world regression datasets. Resampling is a critical technique utilized in imbalanced regression, which simply and directly manipulates the quantity of skewed samples to rebalance the data distribution. However, the relevant academic research mainly focused on classification, with little discussion on imbalanced regression. Therefore, an adaptive Nonlinear Oversampling within Informed Granular-ball Space (NO-IGS) is proposed for imbalanced regression as a novel direction. First, NO-IGS proposes a nonconvex optimization model to fit the distribution of the dataset and adaptively split the granular-ball space. Then, NO-IGS proposes another non-convex optimization model to select an appropriate granular-ball space for oversampling. The existence of optimal solutions for both optimization models is proven, and solution algorithms are provided. Finally, based on anisotropic multivariate Gaussian distributions, new samples conforming to the original distribution of the dataset are synthesized in the granular-ball space. Numerical experiments based on dozens of datasets and comparison algorithms demonstrate the effectiveness and robustness of NO-IGS.