Enhanced tensor low-rank representation learning for multi-view clustering.

Area of Science:

• Machine Learning
• Data Mining
• Computer Vision

Background:

• Multi-view subspace clustering (MSC) methods leverage latent subspace structures in multi-view data.
• Tensor nuclear norm (TNN)-based subspace clustering shows promise but often struggles with intrinsic cluster structure and high-order correlations.
• Existing TNN methods exhibit limitations in exploiting comprehensive data relationships, leading to suboptimal clustering performance.

Purpose of the Study:

• To propose a novel tensor low-rank representation (TLRR) learning method for enhanced multi-view clustering.
• To address the limitations of existing TNN-based methods in capturing intrinsic cluster structures and high-order correlations.
• To develop an effective optimization algorithm for the proposed TLRR method with theoretical convergence guarantees.

Main Methods:

• Construct a 3rd-order tensor from multi-view data features.
• Utilize the t-product in tensor space for self-representation tensor computation.
• Employ the ℓ1,2 norm for class-specific distribution and a weighted TNN for tighter rank approximation on a rotated tensor.

Main Results:

• The proposed TLRR method effectively captures intrinsic cluster structures and high-order correlations.
• An efficient optimization algorithm with proven convergence and low computational complexity was developed.
• Extensive experiments on four datasets demonstrate superior performance compared to state-of-the-art MSC methods.

Conclusions:

• The novel TLRR method significantly advances multi-view subspace clustering.
• TLRR offers improved performance by effectively exploiting tensor-based representations and constraints.
• The proposed approach provides a robust and efficient solution for complex multi-view clustering tasks.