Dual Contrast-Driven Deep Multi-View Clustering.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Science

Background:

• Consensus representation learning is crucial for multi-view clustering.
• Existing methods often fail to learn discriminative representations due to inadequate cluster separation and compactness.

Purpose of the Study:

• To propose a novel deep multi-view clustering network that learns clustering-friendly representations.
• To address limitations in existing methods by incorporating a dual contrastive mechanism.

Main Methods:

• Developed a deep multi-view clustering network utilizing dual contrastive losses: dynamic cluster diffusion loss and reliable neighbor-guided positive alignment loss.
• Employed view-specific encoders for feature extraction and an adaptive feature fusion strategy for consensus representations.
• Implemented a dynamic cluster diffusion module for inter-cluster separation and a neighbor-guided alignment module for within-cluster compactness.

Main Results:

• The proposed method successfully learns representations with strong inter-cluster separation and within-cluster compactness.
• Experimental results demonstrate superior clustering performance compared to state-of-the-art approaches on multiple datasets.

Conclusions:

• The dual contrastive mechanism effectively enhances the quality of learned representations for multi-view clustering.
• The proposed network offers a promising advancement in achieving accurate and robust multi-view clustering.