Propagating Cross-View Semantics for Multi-view Clustering: A Unified Anchor Refinement Paradigm

Area of Science:

• Machine Learning
• Data Mining
• Computer Science

Background:

• Anchor-based methods are successful in large-scale multi-view clustering.
• Existing methods are sensitive to initial anchor quality and have limited multi-view interactions, causing semantic misalignment.

Purpose of the Study:

• To propose a unified framework, Cross-View Semantic Propagation (CVSP), to improve anchor quality and multi-view interactions in clustering.
• To address the limitations of sensitivity to initial anchors and semantic misalignment in existing anchor-based approaches.

Main Methods:

• Developed a plug-and-play framework, CVSP, with two variants: CVSP-F (fixed anchors) and CVSP-U (learnable anchors).
• Implemented an anchor alignment module for cross-view consistency, an anchor refinement module using a cross-view graph for optimization, and a revised anchor graph construction module for robustness.

Main Results:

• Both CVSP-F and CVSP-U consistently outperformed twelve state-of-the-art anchor-based methods across multiple datasets.
• The anchor refinement module demonstrated generalization capabilities when integrated with various existing methods.

Conclusions:

• CVSP offers a robust and effective solution for improving anchor quality and multi-view interactions in clustering.
• The proposed framework enhances clustering performance and robustness, showing significant advantages over existing anchor-based techniques.