Joint diversity regularization and graph regularization for multiple kernel k-means clustering via latent variables

Abstract

Multiple kernel k-means (MKKM) clustering algorithm is widely used in many machine learning and computer vision tasks. This algorithm improves clustering performance by extending the traditional kernel k-means (KKM) clustering algorithm to a multiple setting by combining a group of pre-specified kernels. In this paper, we develop and propose a multiple kernel k-means clustering via latent variables (MKKLV) algorithm, in which base kernels can be adaptively adjusted with respect to each sample. To improve the effectiveness of the kernel-specific and sample-specific characteristics of the data, joint diversity regularization and graph regularization are utilized in the MKKLV algorithm. An efficient three-step iterative algorithm is employed to jointly optimize the kernel-specific and sample-specific coefficients. Experiments validate that our algorithm outperforms state-of-the-art techniques on several different benchmark datasets.