Balanced Multi-view Clustering.

Area of Science:

• Machine Learning
• Data Science
• Computer Vision

Background:

• Multi-view clustering (MvC) integrates information from diverse data views to improve clustering accuracy.
• Existing joint training paradigms in MvC can lead to under-optimized view-specific features due to imbalanced learning objectives.
• Discriminative views may dominate the learning process, hindering the full utilization of multi-view information.

Purpose of the Study:

• To analyze the imbalanced optimization phenomenon in joint-training MvC from a gradient descent perspective.
• To propose a novel balanced multi-view clustering (BMvC) method to overcome limitations of current MvC approaches.
• To enhance the learning of view-specific feature extractors and achieve a better balance between view-specific and view-invariant patterns.

Main Methods:

• Developed a balanced multi-view clustering (BMvC) method incorporating view-specific contrastive regularization (VCR).
• VCR preserves sample similarities from joint and view-specific features within clustering distributions.
• Analysis demonstrates VCR adaptively modulates gradient magnitudes for balanced optimization of view-specific feature extractors.

Main Results:

• The proposed BMvC method effectively balances the exploitation of view-specific patterns and exploration of view-invariant patterns.
• Experiments on eight benchmark MvC datasets and two spatially resolved transcriptomics datasets show superior performance compared to state-of-the-art methods.
• The method demonstrates enhanced capability in capturing underlying data structures by fully leveraging multi-view information.

Conclusions:

• BMvC offers a significant improvement over existing MvC techniques by addressing imbalanced view optimization.
• The VCR mechanism provides adaptive modulation, leading to more effective integration of multi-view data.
• The approach shows promise for applications requiring robust clustering of complex, multi-modal data.