Complementary Learning and Smooth Representation for Multi-view Subspace Clustering

Recent multi-view clustering methods have made significant progress in exploring the consistent and inconsistent features of multi-view data. However, existing methods either discard inconsistent features or directly fuse inconsistent features that contain corruptions, failing in making full utilization of complementary features contained in inconsistency. In addition, many methods fail to maintain the local neighborhood relationship between data in subspace learning, which inevitably loses information. To address these issues, we propose Complementary Learning and Smooth Representation for Multi-view Subspace Clustering (CLSRMC). Firstly, we differentiate the learned inconsistent features into complementary features beneficial for clustering and unfavorable corruptions, and impose different regularization constraints respectively. Secondly, we apply smooth regularization to the complementary features to supplement local structure information. Finally, a consensus graph containing complementary information of each view is obtained through graph fusion. An efficient iterative algorithm is designed to solve the optimization problem of CLSRMC. Extensive experiments on five real-world datasets validate the superiority of our method.