An Improved Ultra-Scalable Spectral Clustering Assessment with Isolation Kernel

Spectral clustering is a well-established unsupervised learning technique that can discover clusters with complex shapes in a dataset. However, applying spectral clustering to large-scale data is often prohibitive due to its high computational cost. To address this issue, the ultra-scalable spectral clustering (U-SPEC) algorithm was recently proposed. In this paper, we improve the performance of U-SPEC by incorporating a data-dependent kernel method. We introduce the Isolation Kernel into the U-SPEC framework, resulting in a novel algorithm called IK-USPEC, which can handle datasets with heterogeneous densities. We evaluate IK-USPEC on 11 real-world and synthetic datasets, and show that it outperforms existing state-of-the-art clustering algorithms.