A Dual-Type Granular Ball Generation Method Based on Multi-Granularity Neighborhood

Granular computing (GrC) significantly improve the efficiency of core algorithms like clustering and classification by reducing data volume. However, existing granular ball generation algorithms have problems such as relying on a single metric and fail to capture both local density homogeneity and neighborhood connectivity. This paper proposes a new algorithm that fuses multi-granularity density and neighborhood relationship, introducing Multi-Granularity Neighborhood Density (MGND) and Multi-Granularity Neighborhood Relationship Degree (MNRD), and classifies granular balls into core and non-core ones accordingly. The algorithm undergoes sequential steps including data preprocessing, initial granular ball generation, splitting optimization, and classification, which enables it to adaptively generate granular balls and improve the ability to represent different data regions. Results show the algorithm adaptively fits data distributions, effectively distinguishes core/non-core regions, and enhances heterogeneous data representation.