Partition-level fusion induced multi-view Subspace Clustering with Tensorial Geman Rank.

Area of Science:

• Machine Learning
• Data Mining
• Computer Vision

Background:

• Tensor-based multi-view clustering captures high-order correlations but suffers from inaccurate tensor rank approximation and noise-sensitive affinity matrix fusion.
• Existing methods may lead to undesired low-rank structures and sub-optimal clustering due to these limitations.

Purpose of the Study:

• To propose a novel multi-view subspace clustering algorithm, PFMSC-TGR, that addresses the limitations of existing tensor-based approaches.
• To enhance the discriminative power of the representation tensor and improve the robustness of information fusion.

Main Methods:

• Introduced Tensorial Geman Rank (TGR) as a tighter surrogate for tensor rank approximation, penalizing singular values for a more discriminative tensor.
• Implemented partition-level fusion to create a consistent indicator matrix, enhancing stability against noisy data.
• Developed a unified framework combining TGR and partition-level fusion, optimized by an efficient algorithm with proven convergence to a stationary KKT point.

Main Results:

• Extensive experiments on nine diverse datasets demonstrated the superiority of PFMSC-TGR compared to eleven state-of-the-art algorithms.
• The proposed TGR constraint led to a strongly discriminative representation tensor.
• Partition-level fusion significantly improved model stability and robustness to noise.

Conclusions:

• PFMSC-TGR offers a more effective and robust approach to multi-view subspace clustering.
• The novel TGR and partition-level fusion strategies significantly advance the field of tensor-based clustering.
• The algorithm's performance and convergence properties are validated through comprehensive experiments and theoretical analysis.