Robust Multi-View Subspace Clustering Via Weighted Multi-Kernel Learning and Co-Regularization

Abstract

Using multi-kernel learning to deal with the non-linear relationship of data has become a new research topic in the field of multi-view subspace clustering. However, the existing methods have the following three defects: 1) the simple consensus kernel weighting strategy cannot give full play to the advantages of multiple kernels; 2) they are sensitive to non-Gaussian noise and their learning affinity matrices cannot meet the block diagonal properties required by clustering, resulting in low clustering performance; 3) the complementary feature information between the data of each view cannot be fully mined. In this paper, a novel robust multi-view subspace clustering method is proposed based on weighted multi-kernel learning and co-regularization (WMKMSC). Based on the self-expression learning framework, block diagonal regularizer (BDR), multi-kernel learning strategy and co-regularization are integrated into the proposed model. Especially, as a robust learning method, the mixture correntropy is used to construct a robust multi-kernel weighting strategy, which is helpful to learn the best consensus kernel. Our method is more effective and robust than several of the most advanced methods on five commonly used datasets.