Robust multi-view low-rank embedding clustering

Significant improvements of multi-view subspace clustering have emerged in recent years. However, multi-view data are often lying on high-dimensional space and inevitably corrupted by noise and even outliers, which pose challenges for fully exploiting the intrinsic underlying relevance of multi-view data, as the redundant and corrupted features are highly deceptive. To address the above problems, this paper proposes a robust multi-view low-rank embedding (RMLE) method for clustering. Specifically, RMLE projects each high-dimensional view onto a clean low-rank embedding space without energy loss, such that multiple high-quality candidate affinity graphs are yielded by using self-expressiveness subspace learning. Meanwhile, it integrates the clean complimentary information of multi-view data in semantic space to learn a shared consensus affinity graph. Further, an efficient alternating optimization algorithm is designed to solve our RMLE by the alternating direction method of multipliers. Extensive experiments on four benchmark multi-view datasets demonstrate the performance superiority and advantages of RMLE against many state-of-the-art clustering methods.