Adaptive Transition Probability Matrix Learning for Multiview Spectral Clustering.

Area of Science:

• Machine Learning
• Data Mining
• Computer Science

Background:

• Multiview clustering is crucial for unsupervised learning but often suffers from high computational costs due to self-representation properties.
• Existing methods may not guarantee optimal transition probability matrices by constructing them in separate steps.

Purpose of the Study:

• To propose a unified model, MCA²M, for multiview spectral clustering that directly learns an adaptive transition probability matrix.
• To address the computational inefficiencies and suboptimal matrix learning in current multiview clustering techniques.

Main Methods:

• Developed a one-step strategy under the robust principal component analysis framework to directly learn the transition probability matrix.
• Employed an alternating optimization algorithm based on the alternating direction method of multipliers.
• Ensured nonnegativity and symmetry of the transition probability matrix without postprocessing.

Main Results:

• The proposed MCA²M method directly learns a nonnegative and symmetric transition probability matrix.
• MCA²M demonstrates superior performance compared to state-of-the-art methods on real-world datasets.
• The method achieves higher efficiency by avoiding separate steps for representation and matrix construction.

Conclusions:

• MCA²M offers an efficient and effective solution for multiview spectral clustering.
• The direct learning of an adaptive transition probability matrix is a key advancement.
• The proposed model provides a robust framework for unsupervised multiview data analysis.