This paper highlights that existing multi-kernel clustering algorithms (e.g., multi-kernel K-means) suffer from computational efficiency and robustness issues in complex data distributions. This is because optimization methods rely on inter-point relationships, making it difficult to accurately capture the inherent structure and diversity of the data set. Furthermore, complex interactions among multiple kernels further exacerbate these issues, impacting the ability to cluster data points in high-dimensional space. In this paper, we improve the multi-kernel clustering framework by leveraging granular-ball computing. The core of granular-ball computing is to adaptively fit the data distribution using balls, ranging from coarse to tolerant, to fit the data distribution. Each ball can enclose data points based on a density consistency measure. This ball-based data description improves computational efficiency and robustness against unknown noise. Specifically, based on the granular-ball representation, we propose the granular-ball kernel (GBK) and its corresponding granular-ball multi-kernel K-means framework (GB-MKKM) for efficient clustering. The proposed GB-MKKM framework using granular ball relations in multi-kernel space demonstrates its efficiency and superior clustering performance in empirical evaluations on various clustering tasks.
