2024 | Suyuan Liu, Junpu Zhang, Yi Wen, Xihong Yang, Siwei Wang, Yi Zhang, En Zhu, Chang Tang, Long Zhao, Xinwang Liu
The paper introduces a novel method called Sample-level Cross-view Similarity Learning (SCSL) for handling incomplete multi-view clustering. Incomplete multi-view clustering (IMVC) is a challenging task due to the partial missingness in multi-view data, which traditional clustering methods struggle to address. Existing similarity-based IMVC methods often fail to construct a complete and accurate similarity matrix, leading to anomalous similarity values and poor clustering results.
SCSL addresses these issues by projecting all samples to a common dimension and constructing a complete similarity matrix across views based on both inter-view and intra-view sample relationships. The method ensures the validity of the projection through a simultaneously learning consensus representation and enhances the quality of the similarity matrix using graph Laplacian regularization. This approach avoids the limitations of individual view similarities and provides a more comprehensive representation of sample relationships.
The proposed SCSL method is evaluated on six benchmark datasets, demonstrating its effectiveness in handling incomplete multi-view clustering tasks. Experimental results show that SCSL outperforms existing methods in terms of accuracy, normalized mutual information (NMI), and purity, especially under varying degrees of incompleteness. The method's robustness and stability are further validated through ablation studies and sensitivity analyses.
Overall, SCSL provides a novel and effective solution for incomplete multi-view clustering, making it a valuable contribution to the field.The paper introduces a novel method called Sample-level Cross-view Similarity Learning (SCSL) for handling incomplete multi-view clustering. Incomplete multi-view clustering (IMVC) is a challenging task due to the partial missingness in multi-view data, which traditional clustering methods struggle to address. Existing similarity-based IMVC methods often fail to construct a complete and accurate similarity matrix, leading to anomalous similarity values and poor clustering results.
SCSL addresses these issues by projecting all samples to a common dimension and constructing a complete similarity matrix across views based on both inter-view and intra-view sample relationships. The method ensures the validity of the projection through a simultaneously learning consensus representation and enhances the quality of the similarity matrix using graph Laplacian regularization. This approach avoids the limitations of individual view similarities and provides a more comprehensive representation of sample relationships.
The proposed SCSL method is evaluated on six benchmark datasets, demonstrating its effectiveness in handling incomplete multi-view clustering tasks. Experimental results show that SCSL outperforms existing methods in terms of accuracy, normalized mutual information (NMI), and purity, especially under varying degrees of incompleteness. The method's robustness and stability are further validated through ablation studies and sensitivity analyses.
Overall, SCSL provides a novel and effective solution for incomplete multi-view clustering, making it a valuable contribution to the field.