MambaMVC: Learning state space model enhanced representation for contrastive multi-view clustering

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Science

Background:

• Deep multi-view clustering methods excel at feature extraction and high-dimensional data handling.
• Existing approaches often use single-layer fusion and overlook inter-view dependencies, limiting feature representation.
• These limitations hinder comprehensive information capture and the expressiveness of local features.

Purpose of the Study:

• To propose MambaMVC, a novel state space model-enhanced representation for multi-view clustering.
• To address limitations in feature fusion and inter-view dependency modeling in existing methods.
• To improve the comprehensive capture of global and local information for enhanced clustering.

Main Methods:

• Introduced dual-layer feature fusion: Global Structure-guided Feature Fusion Learning (GSFL) and Global Bi-directional Mamba (GB-Mamba).
• Employed Cross-view Bi-directional Mamba (CB-Mamba) for enhanced inter-view information exchange and local feature refinement.
• Integrated contrastive learning for label consistency and self-supervised learning guided by global target distribution.

Main Results:

• MambaMVC demonstrated superior performance across eight diverse multi-view datasets.
• The dual-layer fusion effectively captured comprehensive global information.
• Enhanced inter-view communication and local feature learning improved clustering accuracy.

Conclusions:

• MambaMVC effectively addresses limitations of single-layer fusion and ignored inter-view dependencies in deep multi-view clustering.
• The proposed model achieves state-of-the-art results by leveraging Mamba modules for advanced feature representation.
• The integration of contrastive and self-supervised learning further boosts clustering effectiveness and robustness.