Multi-view subspace clustering via adaptive graph learning and late fusion alignment.

Area of Science:

• Machine Learning
• Data Mining
• Computer Vision

Background:

• Multi-view subspace clustering methods leverage complementary information from diverse data sources.
• Existing methods often rely on early fusion or single-view analysis, potentially limiting performance.
• Degeneration in clustering occurs when partitions are fused prematurely without fully exploiting inter-sample relationships.

Purpose of the Study:

• To propose a novel multi-view subspace clustering method named AGLLFA.
• To address limitations of early fusion strategies in existing multi-view clustering techniques.
• To enhance clustering performance by adaptively learning graph structures and employing late fusion alignment.

Main Methods:

• Adaptive graph learning for each view to capture sample similarity.
• Spectral embedding learning to explore latent feature spaces across views.
• Late fusion alignment mechanism for optimal clustering partition generation.
• An alternating updating algorithm with proven convergence for optimization.

Main Results:

• The proposed AGLLFA method demonstrates superior performance compared to state-of-the-art methods.
• Experiments on benchmark datasets validate the effectiveness of adaptive graph learning and late fusion.
• The method successfully exploits complementary information across multiple views for improved clustering.

Conclusions:

• AGLLFA offers a robust approach to multi-view subspace clustering.
• The late fusion strategy effectively integrates view-specific information.
• The method provides a significant advancement in leveraging multi-view data for clustering tasks.