Deep Multi-View Contrastive Clustering via Graph Structure Awareness.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Science

Background:

• Multi-view clustering (MVC) is crucial for unsupervised learning, aiming to uncover relationships in heterogeneous data.
• Existing deep MVC methods often focus solely on attribute features, potentially overlooking valuable structural information.
• The need for methods that effectively integrate view-specific features and preserve inter-view structural information is evident.

Purpose of the Study:

• To propose a novel deep multi-view contrastive clustering method with graph structure awareness (DMvCGSA).
• To enhance the discriminative power of view-specific features by preserving latent structural information.
• To improve clustering performance by directly exploring clustering-beneficial consistency information at the cluster level.

Main Methods:

• Developed a GCN-embedded autoencoder to extract view-specific features while preserving latent structural information.
• Implemented a similarity-guided instance-level contrastive learning scheme to enhance feature distinctiveness.
• Employed cluster-level contrastive learning to capture clustering-relevant consistency information.

Main Results:

• The proposed DMvCGSA method demonstrates superior performance compared to state-of-the-art models.
• Experimental results on twelve benchmark datasets validate the effectiveness of the approach.
• The method successfully exploits collaborative representations through instance- and cluster-level contrastive learning.

Conclusions:

• DMvCGSA offers an effective approach for multi-view clustering by integrating graph structure awareness and contrastive learning.
• The method's ability to preserve structural information and focus on clustering-beneficial consistency is key to its success.
• This work advances the field of unsupervised multi-view learning with a robust and high-performing clustering technique.