TCH: A novel multi-view dimensionality reduction method based on triple contrastive heads.

Area of Science:

• Machine Learning
• Data Science
• Computer Vision

Background:

• Multi-view dimensionality reduction (MvDR) addresses high-dimensional challenges in multi-view data.
• Contrastive learning (CL) shows superior performance but often extracts redundant information and misses view-specific details.

Purpose of the Study:

• To develop an MvDR method that eliminates redundant information.
• To capture view-specific discriminative information effectively.
• To enhance the performance of MvDR by leveraging contrastive learning principles.

Main Methods:

• Proposed a novel MvDR method utilizing triple contrastive heads (TCH).
• Introduced feature- and recovery-level contrastive losses to refine information extraction.
• Integrated sample-, feature-, and recovery-level contrastive losses guided by the information bottleneck principle.

Main Results:

• The TCH method successfully eliminates redundant information.
• View-specific discriminative information is effectively captured.
• Experimental results on five real-world datasets demonstrate superior performance compared to existing methods.

Conclusions:

• The proposed TCH method offers an effective approach for multi-view dimensionality reduction.
• The method aligns with the information bottleneck principle, extracting minimal yet sufficient discriminative information.
• Theoretical insights into the relationship between TCH and mutual information support the method's efficacy.