Consistent learning for incomplete multi-view clustering

Abstract

With the rapid development of artificial intelligence in multi-view data analysis, clustering incomplete multi-view data has become a hot direction for studying data deficiency in real scenarios. Numerous innovative methods have been proposed to tackle these problems effectively. However, most studies ignore the fact that samples with different numbers of available instances should have different weights. In addition, as the structure of multi-view data becomes increasingly complex, we should consider introducing a hyper-Laplacian matrix rather than the traditional Laplacian matrix to mine higher-order semantic information. This paper proposes an effective multi-view clustering approach to address the issues above. We incorporate weight vectors reflecting the number of available instances into the learning process of the consensus representation matrix. In addition, we consider using hyper-Laplacian matrix coupling to represent the structural information of the matrix and the sample. This paper conducts many experiments on eight different datasets and selects nine advanced incomplete multi-view clustering methods for comparison. A large number of experiments demonstrate the effectiveness of our method on incomplete multi-view clustering.