Learning Common Semantics via Optimal Transport for Contrastive Multi-View Clustering.

Area of Science:

• Machine Learning
• Data Science
• Computer Vision

Background:

• Multi-view clustering methods leverage contrastive learning to address representation gaps between data views.
• Existing approaches often fail to align semantics globally, potentially weakening underlying semantic patterns.
• A unified semantic understanding across all views is crucial for robust multi-view clustering.

Purpose of the Study:

• To introduce Common Semantics via Optimal Transport (CSOT), a novel method for enhancing contrastive multi-view clustering.
• To develop a framework that integrates multiple data views into a common semantic space.
• To improve the learning of consistent and discriminative representations from multi-view data.

Main Methods:

• CSOT utilizes optimal transport to map samples from multiple views into joint clusters, representing global multi-view semantic patterns.
• A semantic learning module employs soft assignment vectors from optimal transport as global supervision for consistent semantic learning.
• A semantic-aware re-weighting strategy prioritizes samples based on their semantic significance for improved contrastive learning.

Main Results:

• CSOT effectively integrates information from multiple views into a common semantic space.
• The semantic learning module and re-weighting strategy enhance the consistency of learned semantics across views.
• Experimental results confirm that CSOT achieves state-of-the-art performance in multi-view clustering tasks.

Conclusions:

• CSOT offers a novel approach to multi-view clustering by incorporating global semantic alignment through optimal transport.
• The method successfully addresses the limitations of previous techniques in capturing comprehensive multi-view semantics.
• CSOT demonstrates superior performance, establishing a new benchmark for multi-view clustering.