Label completion based concept factorization for incomplete multi-view clustering

• Published in: Knowledge-based systems Vol. 310; p. 112953
• Main Authors: Yang, Beihua, Song, Peng, Cheng, Yuanbo, Liu, Zhaowei, Yu, Yanwei
• Format: Journal Article
• Language: English
• Published: Elsevier B.V 15.02.2025
• Subjects: Concept factorization; Incomplete multi-view clustering; Spectral rotation; Weighted tensor Schatten p-norm; Concept factorization; Spectral rotation; Incomplete multi-view clustering; Weighted tensor Schatten p-norm
• ISSN: 0950-7051
• DOI: 10.1016/j.knosys.2025.112953

Incomplete multi-view clustering (IMVC) has attracted much attention due to its superior performance in handling incomplete multi-view data. However, existing IMVC methods pay little attention to the semantic associations between incomplete data and concepts. On the other hand, the acquisition of cluster labels also needs to be achieved by clustering algorithms, which splits concept factorization and label learning into two steps. To address these limitations, we design a new IMVC model called label completion based concept factorization (LCCF). Specifically, we first integrate the concept factorization and label learning into the IMVC framework, which can explore the semantic associations between incomplete data and concepts and simultaneously reduce the cost of the completion process. Meanwhile, the weighted spectral rotation is employed to adaptively perform view indicator matrix fusion, which can seamlessly obtain the categories of all samples. Furthermore, we introduce the weighted tensor Schatten p-norm (WTSN) regularization, which can better approximate the rank and exploit the salient structural information in the matrix based on the differences between the singular values. To evaluate the effectiveness of our method, we conduct comprehensive experiments by comparing it with eight baseline methods utilizing five evaluation metrics. The results demonstrate that the proposed LCCF model exhibits superior performance compared to existing state-of-the-art methods. In particular, on the NGs dataset with a missing rate of 50%, the LCCF model exhibits much better performance in terms of ACC and MNI metrics, with an improvement of 8.4% and 22.7%, respectively, in comparison to the second-best algorithm.