Anchors Crash Tensor: Efficient and Scalable Tensorial Multi-View Subspace Clustering.

Area of Science:

• Machine Learning
• Data Science
• Computer Vision

Background:

• Tensorial Multi-view Clustering (TMC) uses low-rank tensor learning for multi-view clustering but faces scalability and accuracy challenges.
• Existing TMC methods struggle with large datasets due to computational complexity and estimation bias from convex rank surrogates.
• A key limitation is the inadequate balance between learning view consistency and complementarity.

Purpose of the Study:

• To propose an efficient and scalable framework for Tensorial Multi-view Clustering (TMC) suitable for large datasets.
• To introduce novel models that address the estimation bias and balance consistency/complementarity in multi-view clustering.
• To develop efficient optimization algorithms for the proposed clustering methods.

Main Methods:

• Developed Efficient and Scalable Tensorial Multi-View Subspace Clustering (ESTMC) integrating anchor representation and non-convex tensor learning with Generalized Non-convex Tensor Rank (GNTR).
• Introduced ESTMC-C, incorporating Enhanced Tensor Rank (ETR), Consistent Geometric Regularization (CGR), and Tensorial Exclusive Regularization (TER) for balanced view representation.
• Designed efficient iterative optimization algorithms with theoretical convergence guarantees for both ESTMC and ESTMC-C.

Main Results:

• The proposed ESTMC framework significantly enhances efficiency for large-scale multi-view clustering.
• ESTMC-C effectively balances consistency and complementarity, yielding improved clustering representations.
• Extensive experiments show the superiority of ESTMC and ESTMC-C over state-of-the-art multi-view clustering algorithms.

Conclusions:

• ESTMC and ESTMC-C offer efficient, scalable, and effective solutions for multi-view clustering problems.
• The novel regularization techniques and non-convex rank approximations improve clustering accuracy and representation.
• These advancements provide a robust approach for handling complex multi-view data in large-scale applications.