Multiview Subspace Clustering via Co-Training Robust Data Representation.

Area of Science:

• Machine Learning
• Data Mining
• Computer Vision

Background:

• Multiview subspace clustering (MSC) algorithms seek consensus reconstruction matrices by leveraging complementary information across multiple data views.
• Existing MSC methods often process raw data or kernel matrices, overlooking potential feature redundancy and arbitrary design.
• This redundancy can lead to suboptimal clustering performance due to noise and irrelevant information.

Purpose of the Study:

• To propose a novel MSC algorithm that simultaneously addresses sample grouping and data redundancy reduction.
• To enhance the discriminative power of data representations for improved clustering accuracy.
• To develop an efficient and robust algorithm for multiview subspace clustering.

Main Methods:

• The proposed algorithm employs eigendecomposition to derive a low-redundancy data representation.
• A unified model integrates sample clustering and redundancy removal, creating a feedback loop where clustering guides representation refinement.
• An alternating and convergent optimization strategy is utilized to solve the underlying mathematical problem.

Main Results:

• The algorithm demonstrates superior performance compared to existing methods on eight benchmark datasets.
• Experimental validation confirms the algorithm's effectiveness, computational efficiency, and robustness against noise.
• The concurrent approach of grouping samples and removing redundancy leads to significant improvements in clustering outcomes.

Conclusions:

• The proposed MSC algorithm effectively handles data redundancy, leading to more accurate and robust clustering.
• The unified model and feedback mechanism enhance the quality of data representation for better discriminative insights.
• The method offers a computationally efficient and noise-resilient solution for multiview subspace clustering tasks.