Robust Subspace Clustering for Multi-View Data by Exploiting Correlation Consensus.

Area of Science:

• Machine Learning
• Data Science
• Computer Vision

Background:

• Multimedia data often comprises multiple views (e.g., color, shape).
• Multi-view data offers complementary information, improving clustering over single-view approaches.
• Existing methods struggle with noise corruption in multi-view data.

Purpose of the Study:

• To develop a subspace clustering method for multi-view data that exploits inter-view correlations.
• To effectively handle noise corruption in multi-view datasets.
• To improve clustering performance by maintaining view encapsulation.

Main Methods:

• A novel objective function with an angular-based regularizer is proposed to obtain multiple sparse representations for each data object.
• A sparsity-based approach refines angular-based data correlations to generate an ideal similarity matrix.
• Spectral clustering is applied to the refined similarity matrix for final subspace clustering.

Main Results:

• The proposed method effectively reaches data correlation consensus across all views.
• The sparsity-based refinement significantly mitigates the impact of noise corruption.
• Experimental results demonstrate the superior effectiveness of the proposed approach for multi-view subspace clustering.

Conclusions:

• The developed subspace clustering technique robustly handles noisy multi-view data.
• The method successfully leverages complementary information from multiple views for enhanced clustering.
• This approach offers a promising solution for analyzing complex multi-view datasets.