Tensorized Bipartite Graph Learning for Multi-View Clustering.

Area of Science:

• Data Science
• Machine Learning
• Computer Vision

Background:

• Existing graph-based multi-view clustering methods face challenges with computational complexity and fail to simultaneously consider inter-view and intra-view similarities.
• Graph construction and eigen-decomposition in current methods lead to significant time burdens.
• A gap exists in methods that effectively leverage both spatial structure and complementary information across multiple views.

Purpose of the Study:

• To propose a novel Tensorized Bipartite Graph Learning (TBGL) framework for multi-view clustering.
• To address the limitations of existing methods by simultaneously considering inter-view and intra-view similarities.
• To develop an efficient and time-economical algorithm for TBGL with guaranteed convergence.

Main Methods:

• A variance-based de-correlation anchor selection strategy is employed for bipartite graph construction.
• TBGL utilizes tensor Schatten p-norm minimization to exploit inter-view similarity and complementary information.
• Intra-view similarity is captured through L1-norm minimization and connectivity constraints on bipartite graphs.

Main Results:

• The proposed TBGL method demonstrates superior performance compared to state-of-the-art multi-view clustering techniques.
• The learned graphs effectively encode discriminative information and directly indicate data clusters via connected components.
• The developed algorithm for TBGL is efficient, time-economical, and exhibits good convergence properties.

Conclusions:

• TBGL offers an effective solution for multi-view clustering by simultaneously addressing inter-view and intra-view similarities.
• The method achieves high clustering performance while being computationally efficient.
• The findings suggest TBGL as a promising advancement in multi-view clustering research.