Transport Barycenter-Guided Sample-to-Cluster Matching for Unaligned Multi-view Clustering

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Mining

Background:

• Multi-view clustering (MVC) methods struggle with real-world data due to semantic and instance-level view-unaligned problems.
• Existing techniques often fail with discrete, non-convex data structures and unreliable cross-view correspondences, weakening cluster boundaries.

Purpose of the Study:

• To develop a novel approach for multi-view clustering that addresses the limitations of current consistency alignment methods.
• To improve the accuracy and robustness of clustering by preserving inter-cluster semantic boundaries and handling discrete cluster manifolds.

Main Methods:

• Introduced Multi-view Discrete Optimal Transport (MvDOT), a cluster-level transport matching framework.
• MvDOT utilizes semi-discrete optimal transport (OT) to learn a global OT barycenter and transports samples to consensus clusters under consistency constraints.

Main Results:

• MvDOT effectively achieves cross-view consistent alignment even with discrete cluster manifolds and unreliable sample correspondences.
• The proposed method preserves inter-cluster semantic boundaries, uncovering underlying cluster structures more accurately.

Conclusions:

• MvDOT offers a robust solution for complex multi-view clustering tasks, outperforming existing methods.
• The framework's ability to handle discrete structures and unreliable correspondences makes it suitable for diverse real-world applications.