Anchor-Sharing and Cluster-Wise Contrastive Network for Multiview Representation Learning.

Area of Science:

• Machine Learning
• Data Science
• Computer Vision

Background:

• Multiview clustering (MVC) partitions samples using unsupervised learning.
• Existing deep clustering methods face challenges with conflicting objectives (reconstruction vs. view consistency) and sample correlations.
• Current contrastive learning (CL) in MVC may create false negative pairs by ignoring cluster information.

Purpose of the Study:

• To propose a novel multiview representation learning network addressing limitations in existing deep clustering and contrastive learning methods.
• To enhance the discriminative power of consensus representations in multiview data.
• To improve the accuracy and robustness of multiview clustering.

Main Methods:

• Developed an anchor-sharing and cluster-wise contrastive learning (CwCL) network.
• Separated view-specific and view-common learning into distinct network branches.
• Introduced an anchor-sharing feature aggregation (ASFA) module and a cluster-wise CL (CwCL) module incorporating transition probabilities.

Main Results:

• The proposed network effectively addresses the conflict between reconstruction and view consistency objectives.
• The ASFA module enhances the discriminative power of common representations by leveraging sample-anchor relationships.
• The CwCL module mitigates issues with false negative pairs in contrastive learning.
• Experimental results show superior performance compared to state-of-the-art methods.

Conclusions:

• The proposed CwCL network offers a significant advancement in multiview representation learning and clustering.
• The method provides a robust framework for learning discriminative and consistent representations from multiple views.
• This approach holds promise for various applications requiring effective unsupervised multiview data analysis.