Tri-Level Consistency-Diversity Calibration for Multi-View Representation Learning.

Area of Science:

• Machine Learning
• Data Science
• Artificial Intelligence

Background:

• Robust representation learning from multi-source data is challenging.
• Existing methods lack fine-grained feature consistency and hierarchical collaboration.
• This limits the precision of learned representations.

Purpose of the Study:

• To propose a novel hierarchical framework for multi-source representation learning.
• To address limitations in existing class-level and instance-level alignment methods.
• To enhance information flow and semantic integrity across different levels of data representation.

Main Methods:

• Introduced the Tri-Level Consistency-Diversity Calibration (TCDC) method.
• Implemented feature-level variance-covariance constraints for fine-grained feature alignment.
• Integrated semantics-guided multi-objective graph learning with contrastive learning at the instance level.
• Utilized class-level attraction-repulsion constraints with category prototypes for enhanced separability.

Main Results:

• TCDC effectively optimizes information flow across feature, instance, and class levels.
• The method demonstrated improved representation precision in extensive experiments.
• Experiments were conducted on multiple public datasets, validating TCDC's effectiveness.

Conclusions:

• TCDC offers a hierarchical approach to multi-source representation learning.
• The proposed method successfully balances consistency and diversity across multiple levels.
• TCDC significantly advances the state-of-the-art in representation learning from multi-source data.