Sequential multi-view subspace clustering.

Area of Science:

• Computer Science
• Machine Learning
• Data Mining

Background:

• Existing self-representation subspace learning methods often overlook inter-view differences, hindering accurate clustering.
• Current approaches may use inflexible tensor norms, limiting practical applicability.

Purpose of the Study:

• To develop a tensorized multi-view subspace clustering algorithm that addresses limitations of existing methods.
• To enhance the characterization of clustering structures by considering multi-view data differences.

Main Methods:

• Matrix factorization to decompose self-representation matrices into orthogonal projection and affinity matrices.
• Incorporation of L1,2-norm regularization on affinity representation for improved cluster structure depiction.
• Utilization of weighted tensor Schatten p-norm for capturing higher-order structures and complementary multi-view information.

Main Results:

• The proposed method automatically assigns optimal weights to different views without extra parameters.
• An adaptive loss function enhances robustness to outliers and efficient data distribution learning.
• Experimental results demonstrate superior performance compared to state-of-the-art multi-view subspace clustering techniques.

Conclusions:

• The developed tensorized multi-view subspace clustering method offers significant improvements in clustering performance.
• The approach effectively leverages multi-view data by capturing higher-order structures and complementary information.
• The method shows robustness and efficiency, outperforming existing techniques on various datasets.