A1618
Title: Asymptotic behaviors of kernel hierarchical clustering under high-dimensional settings
Authors: Kento Egashira - Tokyo University of Science (Japan) [presenting]
Yota Takao - Tokyo University of Science (Japan)
Abstract: In high-dimensional settings with multiple underlying populations, the asymptotic behavior of hierarchical clustering methods, including Ward's method, has been extensively studied. Under suitable conditions, the resulting dendrogram converges to one of several distinct topological forms, which enables a precise characterization of when the true population structure is correctly recovered. Despite the widespread use and empirical success of kernel methods in high-dimensional data analysis, the theoretical properties of hierarchical clustering combined with kernel techniques remain largely unexplored. The asymptotic behavior of kernel-based hierarchical clustering is investigated to relax existing conditions required for consistent population separation. Focusing on Ward's method with a Gaussian kernel, sufficient conditions are derived under which the desired dendrogram structure is asymptotically obtained. The theoretical analysis provides an explanation for the improved clustering performance observed in high-dimensional settings and broadens the applicability of hierarchical clustering methods.
