Granular-Ball Regeneration Clustering With Principle of Justifiable Granularity

Classical clustering algorithms such as k-means face limitations in handling clusters with heterogeneous shapes, densities, and sizes, while exhibiting sensitivity to initial centroid selection. To overcome these challenges, this article proposes a novel clustering framework based on regenerated granular ball (RGGB) with the principle of justifiable granularity. Unlike existing granular-ball (GB) techniques that overemphasize purity criteria at the expense of uncontrolled ball sizes, RGGB dynamically adjusts granularity levels through iterative regeneration, achieving an optimal balance between detailed data representation and computational efficiency. This adaptability enhances stability in capturing data similarities while mitigating sensitivity to initialization. To validate the method, we integrate RGGB with a novel k-nearest neighbor (KNN) classifier using regenerated GBs to evaluate classification performance and demonstrate practical applications. Experiments on diverse public and realistic datasets demonstrate that the RGGB-based KNN algorithm consistently outperforms existing techniques, including traditional KNN and other methods, making a promising advancement in clustering and classification tasks.