Self-weighted dual contrastive multi-view clustering network.

Area of Science:

• Machine Learning
• Data Mining
• Computer Vision

Background:

• Multi-view Clustering (MVC) leverages consensus from multiple data perspectives.
• Traditional MVC methods struggle with representation degeneration and poor cluster separability.
• Existing approaches often neglect the crucial aspect of inter-cluster distinctiveness.

Purpose of the Study:

• To propose a novel deep multi-view clustering network addressing representation degeneration and enhancing cluster separability.
• To develop a method that learns discriminative representations with clustering-friendly structures.
• To improve the performance and structure of multi-view clustering.

Main Methods:

• Utilized view-specific autoencoders for latent feature extraction.
• Implemented global feature fusion for consensus information across views.
• Introduced an adaptive weighted mechanism to manage view reliability during fusion.
• Developed a Dynamic Cluster Diffusion (DC) module within a contrastive learning framework to maximize inter-cluster distances.

Main Results:

• Achieved state-of-the-art clustering performance across multiple datasets.
• Demonstrated significant improvement in the separability of learned clusters.
• Effectively mitigated representation degeneration issues through adaptive view weighting.
• Successfully learned clustering-friendly discriminative representations.

Conclusions:

• The proposed Contrastive Learning-based Dual Contrast Mechanism Deep Multi-view Clustering Network offers a robust solution for MVC.
• The method enhances both clustering accuracy and the structural quality of the learned representations.
• The Dynamic Cluster Diffusion module is key to improving inter-cluster separability and overall performance.