A Novel Adaptive Hyperspherical Oversampling Method Based on Extended Natural Neighborhood for Imbalanced Classification

Classifying imbalanced datasets remains a significant challenge for classifiers, with oversampling techniques being a widely used solution. However, many existing oversampling methods are susceptible to noise points, outliers, and related hyperparameters sensitivity, which can degrade their effectiveness. To address these limitations, this paper proposes a novel oversampling method, the Adaptive Hyperspherical Oversampling Method Based on Extended Natural Neighborhood (AHOBENN). The proposed method begins by partitioning the dataset into regions using the extended natural neighborhood approach. Hyperspheres are then constructed around minority class borderline points to define targeted oversampling regions. By leveraging the law of universal gravitation and the characteristics of the extended natural neighborhood, adaptive sampling weights are assigned to each hypersphere, allowing for parameter-free oversampling. Additionally, the Differential Evolution (DE) algorithm is applied to optimize the positions of noise and outlier points, rather than eliminating them. Extensive experiments were conducted on synthetic and public datasets across four different classifiers. Comparative analysis with nine other oversampling methods demonstrates that the proposed method significantly enhances classification performance on imbalanced datasets.