A multiple-kernel based subspace clustering method

Abstract

Spectral clustering has been successfully used in the domain of pattern recognition and computer vision. Kernel subspace clustering has become a hot research topic because it can reveal the nonlinear structure. However, the performance of exiting single kernel subspace clustering relys heavily on the choice of kernel function. To address the problem, we propose a novel method called multiple-kernel based subspace clustering method (MKSC) by combining kernel block diagonal representation with multiple kernel learning. The proposed MKSC algorithm firstly obtains the optimal kernel matrix by using multiple kernel clustering method, then replace the kernel function in single kernel subspace clustering model with the optimized kernel matrix, finally the clustering result is got by optimizing the MKSC model. Experimental results on three datasets testify the effectiveness of our proposed MKSC method.