A Unified Framework for Pseudo-Supervised Clustering via Weighted Sample Aggregation.

Area of Science:

• Machine Learning
• Data Science
• Computer Vision

Background:

• Multi-view clustering methods using graph learning show promise but often lack effective supervisory signals.
• Existing methods struggle to fully exploit latent correlations in multi-view data and equally treat samples in local learning, limiting structure capture.

Purpose of the Study:

• To propose a Unified Framework for Pseudo-Supervised Clustering via Weighted Sample Aggregation (PSC-WSA).
• To address limitations in existing multi-view clustering by incorporating pseudo-supervision and adaptive sample weighting.

Main Methods:

• PSC-WSA framework integrates pseudo-supervision generation, weighted sample aggregation, clustering, and label propagation.
• Sample aggregation is learnable and interacts synergistically with clustering, guided by pseudo-supervision.
• Two weighted aggregation strategies adaptively model local relationships between similar samples across different views.

Main Results:

• The proposed framework effectively enhances the discriminability of representative samples.
• Experiments demonstrate the feasibility, effectiveness, and strong scalability of PSC-WSA on multi-view and remote sensing datasets.
• The method successfully improves clustering performance by leveraging latent correlations and critical local structures.

Conclusions:

• PSC-WSA offers a robust approach to pseudo-supervised multi-view clustering.
• The adaptive sample aggregation and pseudo-supervision strategies significantly improve clustering accuracy and scalability.
• This framework provides a valuable tool for analyzing complex multi-view and multimodal data.