Tensorized Soft Label Learning Based on Orthogonal NMF.

Area of Science:

• Data Science
• Machine Learning
• Computer Vision

Background:

• Multiview high-dimensional data analysis is crucial across various domains.
• Nonnegative matrix factorization (NMF) is a common technique for clustering such data.
• Existing NMF-based methods often neglect spatial structures and require postprocessing, limiting stability.

Purpose of the Study:

• To develop a more stable and interpretable multiview clustering algorithm.
• To address limitations of existing NMF methods by incorporating spatial information.
• To enable direct cluster label extraction without postprocessing.

Main Methods:

• Proposed a novel method minimizing the Schatten p-norm of a tensor composed of coefficient matrices from different views.
• Incorporated orthogonal constraints on the cluster index matrix for sparsity and interpretability.
• Utilized adaptive weights to differentiate the importance of various views.
• Developed an unsupervised optimization scheme for model solving and analysis.

Main Results:

• The proposed method effectively captures complementary information by considering spatial structures within coefficient matrices.
• Orthogonal constraints led to sparse cluster index matrices, allowing direct label acquisition.
• Adaptive weights improved the handling of varying view importance.
• Empirical evaluations on six benchmark datasets demonstrated superior performance compared to existing algorithms.

Conclusions:

• The novel multiview clustering approach offers improved stability and interpretability.
• Considering spatial structure and employing adaptive weights are key to enhancing NMF-based clustering.
• The method provides a direct and robust way to obtain cluster labels from high-dimensional multiview data.