Self-representation and Class-Specificity Distribution Based Multi-View Clustering

Yu Yun,Wei Xia,Yongqing Zhang,Quanxue Gao,Xinbo Gao
DOI: https://doi.org/10.1016/j.neucom.2021.01.039
IF: 6
2021-01-01
Neurocomputing
Abstract:Despite the promising performance for clustering, weighted tensor nuclear norm based multi-view subspace clustering needs to artificially predefine a weighted vector when shrinking all the singular values of tensor. It is very difficult to select a suitable weighted vector due to the complex and unknown distribution of data in real applications. Another limitation is that, the learned affinity matrix does not well exploit the cluster structure. To tackle these problems, we propose a novel multi-view subspace clustering method, called SCSD-MVC. The method adaptively allocates a suitable weighted vector to all singular values. The weighted vector well characterizes the significant difference between the singular values. Drawing the inspiration of one finding that ℓ1,2-norm can well characterize class-specificity distribution, which well encodes the cluster structure, we impose ℓ1,2-norm constraint on affinity matrix such that it can well characterize the cluster structure. Experimental results on several databases indicate that SCSD-MVC outperforms state-of-the-art self-representation based multi-view subspace clustering methods.
What problem does this paper attempt to address?