Consistency- and Inconsistency-Aware Multi-view Subspace Clustering

Multi-view subspace clustering has emerged as a crucial tool to solve the multi-view clustering problem. However, many of the existing methods merely focus on the consistency issue when learning the multi-view representations, failing to capture the latent inconsistency across different views (which can be caused by the view-specificity or diversity). To tackle this issue, we therefore develop a Consistency- and Inconsistency-aware Multi-view Subspace Clustering for robust clustering. In the proposed method, we decompose the multi-view representations into a view-consistent representation and a set of view-inconsistent representations, through which the multi-view consistency as well as multi-view inconsistency can be well explored. Meanwhile, our method aims to suppress the redundancy and determine the importance of different views by introducing a novel view weighting strategy. Then a unified objective function is constructed, upon which an efficient optimization algorithm based on ADMM is further performed. Additionally, we design a new way to compute the affinity matrix from both consistent and inconsistent perspectives, which makes sure that the learned affinity matrix comprehensively fit the inherent properties of multi-view data. Experimental results on multiple multi-view data sets confirm the superiority of our method.