MC-NCLWO: Newton’s Cooling Law-Based Oversampling Algorithm for Multi-Class Imbalanced Datasets With Feature Noise

Imbalanced data classification remains a fundamental challenge in machine learning, especially in multi-class scenarios where feature noise, class overlap, and small disjunct sub-concepts are prevalent. While various oversampling techniques have been proposed, most existing methods are primarily designed for binary classification and often struggle with noise sensitivity and structural preservation in complex multi-class settings. To address these gaps, this paper proposes a Newton’s Cooling Law-Based Weighted Oversampling strategy (MC-NCLWO) specifically tailored for imbalanced multi-class datasets. First, a dynamic noise-cleaning mechanism inspired by Newton’s Cooling Law is introduced to adaptively identify and eliminate noisy majority class samples based on the classification difficulty of each instance. Second, a difficulty-weighted sample generation strategy is employed to maintain local class balance and preserve the intrinsic structure of minority class sub-concepts. Furthermore, a sampling-priority strategy is designed to coordinate the oversampling process among multiple classes, thereby reducing mutual interference and ensuring the structural integrity of each minority class. Comprehensive experiments on 23 benchmark datasets demonstrate that MC-NCLWO consistently outperforms 10 state-of-the-art methods across multiple performance metrics, including $F_{\beta M}$ , AvAcc, CBA, and CEN. Additionally, the proposed method exhibits strong robustness in high-noise environments, significantly mitigating performance degradation caused by feature-level noise.