Kernel-based clustering via Isolation Distributional Kernel

Abstract

Clustering has become one of the widely used automatic data-labeling techniques applied in a variety of disciplines. Kernel-based clustering is a technique designed to identify non-linearly separable clusters with irregular shapes. However, existing kernel-based clustering algorithms usually produce weaker clustering outcomes than density-based clustering, and they have high computational cost which limits their applications to small datasets only. In this paper, we contend that these limitations are mainly due to the use of (a) the Expectation-and-Maximization algorithm as an optimization procedure, and (b) a non-adaptive kernel. In addition, to address the limitations of current kernel-based algorithms, we propose the first clustering algorithm that employs an adaptive distributional kernel without any optimization, while achieving a similar optimization objective function. We demonstrate its superior performance of identifying complex clusters on massive datasets under different real-world application scenarios.