Structure-Enhanced Self-Supervised Weighted Information Bottleneck for Multiview Clustering.

Area of Science:

• Data Mining and Pattern Recognition
• Machine Learning
• Artificial Intelligence

Background:

• Multiview clustering (MVC) leverages correlations between data views for consistent structure discovery.
• Weighted MVC methods prioritize view quality but often neglect self-supervision and result consistency.
• Existing approaches inadequately consider the interplay between individual view clustering and the final outcome.

Purpose of the Study:

• To propose a novel method, the structure-enhanced self-supervised weighted information bottleneck (S2WIB), for improved multiview clustering.
• To enhance view-weight learning by integrating view-contained and self-supervised information.
• To fully exploit multiview data potential by considering complementary and consistent cluster structure information.

Main Methods:

• Developed a view-weight learning mechanism incorporating view-contained and self-supervised information.
• Integrated weighted information from diverse views using an information bottleneck (IB) framework.
• Explored view correlations via complementary information and consistent view cluster structure.

Main Results:

• The S2WIB method demonstrated effective performance across diverse datasets including text, image, video, multimodal, and multiomics data.
• Experimental results indicate superior performance compared to existing weighted MVC methods.
• The approach successfully leverages both self-supervision and structural consistency for enhanced clustering.

Conclusions:

• The proposed S2WIB method offers a significant advancement in multiview clustering by effectively integrating self-supervision and structural consistency.
• The method's ability to learn robust view weights and exploit comprehensive view correlations leads to superior clustering performance.
• S2WIB shows broad applicability and effectiveness across various data types and domains, highlighting its practical utility.