A hybrid sampling algorithm for imbalanced and class-overlap data based on natural neighbors and density estimation

Imbalanced data classification poses a significant challenge in machine learning and data mining, exacerbated by class overlap which adversely affects model performance. Resampling is widely used to tackle imbalanced data. However, most resampling algorithms overlook the complexity of overlapping region samples, resulting in excessive removal of majority class samples and insufficient representation of minority class information. To tackle these issues, this paper proposes a novel approach, namely natural neighbors and density estimation-based hybrid sampling algorithm (NaNDS). NaNDS fully considers the varying impacts of overlapping samples with different characteristics on subsequent classification tasks. First, the density of minority class samples is estimated using a Gaussian kernel function, and high-density minority class samples are selected to construct a set of hyper-spherical structures. These structures facilitate the geometric identification and removal of overlapping majority class samples that negatively impact classification. Then, oversampling weights for minority samples are determined by combining density and information entropy estimates derived from natural neighbors. Afterward, an adaptive oversampling strategy is developed, using differentiation-based generation guided by natural neighbor information. This process appropriately corrects decision boundary by enhancing information in both overlapping and minority class-dominated regions. Experimental validation on the KEEL dataset demonstrates NaNDS outperforming eight state-of-the-art algorithms, highlighting its superior competitiveness and robustness in handling class-overlap imbalanced datasets.