Manifold-based Incomplete Multi-view Clustering via Bi-Consistency Guidance

Manifold-based Incomplete Multi-view Clustering via Bi-Consistency Guidance

16 May 2024 | Huibing Wang, Mingze Yao*, Yawei Chen, Yunqiu Xu, Haipeng Liu, Wei Jia, Xianping Fu, Yang Wang, Senior Member, IEEE
The paper introduces a novel method called Manifold-based Incomplete Multi-view Clustering via Bi-consistency Guidance (MIMB) to address the challenge of incomplete multi-view clustering. MIMB aims to recover missing instances across multiple views and explore consistent information among them. The method uses a recovery strategy to learn a consensus representation and employs reverse regularization to reduce noise and improve the accuracy of the recovered data. Additionally, MIMB incorporates manifold embedding to capture the local structure of the recovered data, ensuring better clustering performance. The optimization process involves iterative updates of various variables, including the consensus representation, manifold graph, Laplacian representation, basis matrix, recovery matrix, and adaptive weights for each view. Extensive experiments on six benchmark datasets demonstrate that MIMB outperforms several state-of-the-art methods in terms of accuracy, normalized mutual information, and purity, even under high incomplete rates.The paper introduces a novel method called Manifold-based Incomplete Multi-view Clustering via Bi-consistency Guidance (MIMB) to address the challenge of incomplete multi-view clustering. MIMB aims to recover missing instances across multiple views and explore consistent information among them. The method uses a recovery strategy to learn a consensus representation and employs reverse regularization to reduce noise and improve the accuracy of the recovered data. Additionally, MIMB incorporates manifold embedding to capture the local structure of the recovered data, ensuring better clustering performance. The optimization process involves iterative updates of various variables, including the consensus representation, manifold graph, Laplacian representation, basis matrix, recovery matrix, and adaptive weights for each view. Extensive experiments on six benchmark datasets demonstrate that MIMB outperforms several state-of-the-art methods in terms of accuracy, normalized mutual information, and purity, even under high incomplete rates.
Reach us at info@study.space