Generalized Nonconvex Low-Rank Tensor Approximation for Multi-View Subspace Clustering.

Area of Science:

• Machine Learning
• Data Science
• Computer Vision

Background:

• Low-rank tensor representation (LRTR) enhances multi-view clustering by considering data and view relationships.
• Existing LRTR methods use convex approximations (tensor nuclear norm) leading to biased rank estimation.

Purpose of the Study:

• To propose a novel generalized nonconvex low-rank tensor approximation (GNLTA) model for multi-view subspace clustering.
• To address the limitations of convex approximations in LRTR by employing a nonconvex approach.

Main Methods:

• Developed GNLTA to capture high-order correlations among multiple views using low-rank tensor approximation.
• Introduced a generalized nonconvex low-rank tensor norm to better interpret singular values.
• Designed a unified solver for the GNLTA model, proving convergence properties.

Main Results:

• GNLTA effectively captures high-order correlations, surpassing pairwise correlation methods.
• The proposed generalized nonconvex tensor norm provides a more accurate estimation.
• Extensive experiments on seven benchmark datasets show GNLTA outperforms state-of-the-art methods.

Conclusions:

• GNLTA offers a superior approach to multi-view subspace clustering compared to existing LRTR techniques.
• The nonconvex tensor norm is crucial for accurate rank approximation and improved clustering performance.
• The developed solver guarantees convergence to stationary points, ensuring model reliability.