Hierarchical feature based dual contrastive multiview clustering.

Area of Science:

• Machine Learning
• Computer Vision
• Data Science

Background:

• Multiview clustering (MVC) effectively uses information from multiple data sources.
• Existing deep MVC methods often struggle to balance feature consistency and complementarity.
• Suboptimal global representations limit the exploitation of inter-view relationships.

Purpose of the Study:

• To propose a novel deep MVC framework, hierarchical feature based dual contrastive learning (HFDCL).
• To enhance the balance between consistency and complementarity in learned representations.
• To improve the performance of deep multiview clustering.

Main Methods:

• Introduced a hierarchical self-attention-based architecture to disentangle consistency and reconstruction objectives.
• Preserved both localized and holistic features within each view.
• Incorporated a dual contrastive learning module using InfoNCE to maximize mutual information between view-specific and global representations.

Main Results:

• The hierarchical design effectively preserves intrinsic features of each view.
• Dual contrastive learning enhanced feature diversity and consistency across views.
• Extensive experiments demonstrated HFDCL's superior performance over state-of-the-art deep MVC methods.

Conclusions:

• HFDCL offers an effective approach to multiview clustering by optimizing feature representation.
• The proposed framework successfully balances feature consistency and complementarity.
• HFDCL shows significant potential for advancing deep multiview clustering tasks.