非凸张量多视图子空间聚类

Nonconvex tensor multi-view subspace clustering

为了探索非凸方法在多视图聚类方面的应用,基于非凸替换函数和子空间学习,提出非凸张量多视图子空间聚类算法.该算法不仅对多视图数据进行自表示学习来达到学习低维子空间的目的,而且采用带有旋转的张量结构对张量的高阶关联进行挖掘.同时,使用非凸函数替换和广义奇异值算子进行张量最小化问题的求解,从而实现对张量秩的近似.最后基于联合优化所得关联/仿射矩阵实现聚类操作,在不同类型的多视图数据集上的大量实验验证了该方法的聚类效果.

To explore the application of non convex methods for multi-view clustering,this paper proposes a nonconvex tensor multi-view subspace clustering algorithm based on nonconvex substitution functions and subspace learning.The algorithm first performs self-representation attribute learning on multi-view data to achieve dimensionality reduction.Secondly,the tensor structure with rotation is used to mine the spatial association of the tensor.At the same time,the tensor minimization problem is solved using nonconvex function substitution and generalized singular value operator to achieve an approximation of the tensor rank.Finally,the learned correlation matrix is used to implement the clustering operation.Extensive experiments on different types of multi-view datasets validate the clustering effect of the method.