Self-supervised semantic graph propagation for multi-view clustering.

Area of Science:

• Machine Learning
• Data Mining
• Computer Vision

Background:

• Multi-view clustering enhances performance by exploiting data consistency across diverse views.
• Existing methods often neglect pseudo-label graph-based self-supervision, limiting feature discovery in unsupervised settings.

Purpose of the Study:

• To propose a self-supervised semantic graph propagation framework (SemGProp) for effective multi-view clustering.
• To address the limitations of existing methods by incorporating pseudo-label graph-based self-supervision.

Main Methods:

• SemGProp integrates information across views via a fusion mechanism to create a global representation and a pseudo-label graph.
• A consistency graph propagation module refines view-specific similarity graphs using the global pseudo-label structure.
• Kullback-Leibler divergence-based graph loss enforces consistency and aligns feature structures across views.

Main Results:

• The framework effectively regularizes learned representations with reliable class-level information.
• SemGProp ensures consistent structures for semantically similar samples across different views.
• Experimental results demonstrate SemGProp's superiority over state-of-the-art methods in enhancing cross-view structural consistency.

Conclusions:

• SemGProp significantly improves multi-view clustering by effectively utilizing self-supervised semantic graph propagation.
• The proposed method enhances the structural consistency of feature representations across views, leading to better clustering performance.