Uniform Distribution Non-Negative Matrix Factorization for Multiview Clustering

Multiview data processing has attracted sustained attention as it can provide more information for clustering. To integrate this information, one often utilizes the non-negative matrix factorization (NMF) scheme which can reduce the data from different views into the subspace with the same dimension. Motivated by the clustering performance being affected by the distribution of the data in the learned subspace, a tri-factorization-based NMF model with an embedding matrix is proposed in this article. This model tends to generate decompositions with uniform distribution, such that the learned representations are more discriminative. As a result, the obtained consensus matrix can be a better representative of the multiview data in the subspace, leading to higher clustering performance. Also, a new lemma is proposed to provide the formulas about the partial derivation of the trace function with respect to an inner matrix, together with its theoretical proof. Based on this lemma, a gradient-based algorithm is developed to solve the proposed model, and its convergence and computational complexity are analyzed. Experiments on six real-world datasets are performed to show the advantages of the proposed algorithm, with comparison to the existing baseline methods.

Automated summary

This article proposes a tri-factorization-based NMF model with an embedding matrix to generate decompositions with uniform distribution in the learned subspace, making the learned representations more discriminative and leading to higher clustering performance. Additionally, a new lemma is introduced to provide formulas about the partial derivation of the trace function with respect to an inner matrix, along with its theoretical proof, to develop a gradient-based algorithm for solving the proposed model and analyze its convergence and computational complexity. Experiments on six real-world datasets demonstrate the advantages of the proposed algorithm compared to existing baseline methods.