Orthogonal Diversity Nonnegative Matrix Factorization for multi-view clustering

Xinling Zhang, Chengcai Leng, Jinye Peng, Irene Cheng, Anup Basu

Abstract: In the context of rapid development of artificial intelligence, how to extract valuable information from complex multidimensional data has become a core research problem. Multi-view clustering methods based on non-negative matrix factorization (NMF) are widely used in multi-view data analysis, but still face many challenges in practical applications. Current multi-view clustering methods usually solve the problem of diversity among viewpoints by orthogonalization of view representations. However, they fail to fully utilize the rich features of each viewpoint because data from different viewpoints may be interrelated. In addition, existing methods fail to fully consider the orthogonality between base matrices while emphasizing the diversity of view representations. For this reason, this paper proposes a new orthogonal diversity non-negative matrix factorization method (ODNMF). First, ODNMF explores the orthogonality of the representations of sample pairs between different viewpoints. This approach preserves the characteristics of each perspective and enhances the diversity of data representations. Second, ODNMF orthogonalizes the basis matrix of each viewpoint to reduce redundant features and enhance data interpretability and representation. Finally, ODNMF introduces graph regularization for each view to reveal the intrinsic geometric and structural information of features. Experimental results show that ODNMF significantly outperforms existing state-of-the-art algorithms on seven datasets.