Multiview Data Clustering with Similarity Graph Learning Guided Unsupervised Feature Selection.

Area of Science:

• Computer Science
• Data Science
• Machine Learning

Background:

• Multiview data clustering aims to leverage consistent or complementary information across multiple data sources for improved results.
• Challenges in multiview clustering include high dimensionality, lack of labels, and data redundancy, which can negatively impact clustering performance.
• Existing methods often struggle to effectively integrate information from diverse views while addressing these inherent challenges.

Purpose of the Study:

• To develop a novel clustering algorithm, multiview feature selection clustering (MFSC), that addresses the limitations of traditional multiview clustering.
• To combine the strengths of similarity graph learning and unsupervised feature selection for enhanced clustering accuracy.
• To retain essential clustering characteristics while preserving the underlying manifold structure of multiview data.

Main Methods:

• The proposed MFSC algorithm integrates similarity graph learning with unsupervised feature selection.
• Local manifold regularization is incorporated into the similarity graph learning process.
• Clustering labels derived from similarity graph learning serve as the criterion for unsupervised feature selection.

Main Results:

• The MFSC algorithm effectively retains clustering label characteristics while maintaining the manifold structure of multiview data.
• Systematic evaluations were conducted using benchmark multiview datasets and simulated data.
• Experimental results demonstrate that the MFSC algorithm achieves superior performance compared to traditional multiview clustering algorithms.

Conclusions:

• The developed MFSC algorithm offers a robust approach to multiview data clustering.
• The integration of similarity graph learning and unsupervised feature selection proves effective in overcoming common challenges.
• MFSC demonstrates significant improvements in clustering effectiveness over existing methods.