One-step Kernel Multi-view Subspace Clustering

摘要

Multi-view subspace clustering is essential to many scientific problems. However, most existing methods suffer from three aspects of issues. First, these methods usually adopt a two-step framework, lacking the ability to achieve an optimal common affinity matrix across multiple views. Second, these methods are intended to solve the clustering problem in linear subspaces but may fail in practice as most real-world data sets may exhibit non-linear structures. Third, most existing subspace-based methods force the negative elements in the coefficient matrix to be positive, which may damage the inherent correlation among the data. To address above issues, we propose a novel approach termed One-step Kernel Multi-view Subspace Clustering (OKMSC). The common affinity matrix is learned from all views under one-step framework, which integrates the nonnegative and discriminative property of affinity matrix into the computation. Further, a kernelized model is designed to address the nonlinear multi-view clustering problem. And an iterative optimization method is designed to solve the objective function in this model. Extensive experiments have validated the superiority of the proposed method over several state-of-art clustering methods.