One-Stage Multi-view Clustering with Hierarchical Attributes Extraction

Multi-view clustering (MVC) has received significant attention, and obtained praiseworthy performance improvement in comparison with signal-view clustering, since it can effectively take advantage of the underlying correlation and structure information of multi-view data. However, existing methods only utilize signal-layer mapping to exploit clustering information, and ignore the underlying hierarchical attribute information in complex and interleaved multi-view data. In this work, we propose a novel MVC method, one-stage multi-view clustering with hierarchical attributes extracting (OS-HAE), to exploit the underlying hierarchical attributes for MVC. Specifically, we learn multiple latent representations from each view by a novel deep matrix factorization (DMF) framework with a layer-wise scheme, so that the learned representations can contain the hierarchical attribute information of original multi-view data. In addition, the samples from the same clusters but from different views are forced to be closer, and samples from different cluster are away from each other in the latent low-dimensional space. Furthermore, we introduce local manifold learning to guide DMF, such that the deepest representations can preserve structure information of original data. Meanwhile, a novel auto-weighted spectral rotating fusion (ASRF) paradigm is proposed to obtain the final clustering indicator matrix directly, so that OS-HAE can avoid obtaining suboptimal results caused by a two-stage strategy. Then, an alternate algorithm is designed to solve the objective function. Experimental results on six datasets demonstrate the advancement and effectiveness of the proposed OS-HAE. Consequently, the proposed method can effectively exploit the hierarchical information of multi-view to improve clustering performance.