Orthogonal multi-view tensor-based learning for clustering

Abstract

Multi-view spectral clustering aims to improve the performance of spectral clustering through multi-view data. Many multi-view spectral clustering methods have been proposed recently and achieved promising performance. Among these methods, most of them are designed to pursue numerical consistency in multi-view similarity matrices. However, each similarity matrix has its unique statistic distribution, which makes it not appropriate to seek numerical consistency in multi-view similarity matrices or directly average the multi-view similarity matrices. To overcome the aforementioned problem, we propose a novel Orthogonal Multi-view Tensor-based Learning for clustering, abbreviated as OMTL. Specifically, OMTL introduces an orthogonal matrix factorization to eliminate the view-specific statistic distribution and preserve the intrinsic clustering structure of each view, which fully considers the consensus information contained in multiple views to boost multi-view spectral clustering performance. Further, we employ a low-rank tensor constraint to explore the high order correlations among multiple views. By designing an alternating direction method of multipliers (ADMM) based optimization algorithm, the intrinsic similarity matrix of multi-view data can be efficiently learned for spectral clustering. Extensive experiments on several benchmark datasets have illustrated the superior clustering performance of the proposed method compared to several state-of-the-art multi-view clustering methods.