An incomplete multiview clustering approach considering missing data recovery based on consistency.

Abstract

Real-world multiview data often suffers from complex missingness problems, leading to significant performance degradation of clustering methods. Existing methods usually focus only on data completion while ignoring inter-view consistency, or the recovery of missing data is unreliable. For this reason, this paper proposes an incomplete multiview clustering algorithm for recovering missing data based on multiview characteristics. Unlike existing methods, the proposed method achieves reliable recovery of missing data and clustering optimization through consistency preservation. First, a latent subspace representation shared among views is constructed, and the local structure of each view is aligned to the global consensus graph through adaptive graph learning to solve the dimensional heterogeneity problem effectively. Second, the clustering metrics of non-missing samples are used to guide the iterative optimization of missing data to ensure the distributional consistency between the complementary data and the existing instances. Finally, a view weight assignment strategy is introduced to adjust the contribution of each view according to its difference from the consensus graph. The model improves the clustering performance synchronously during the data complementation process. Experiments on multiple datasets show the superior performance of the proposed method over various approaches.