Multiview Clustering via Enhanced Multiorder Bipartite Graph Learning

Area of Science:

• Machine Learning
• Data Mining
• Graph Theory

Background:

• Existing bipartite graph-based multiview clustering (MVC) methods often neglect neighborhood structures and high-order correlations.
• Current approaches overlook relationships among anchor points, limiting the discriminative power of learned graphs.

Purpose of the Study:

• To propose an enhanced multiorder bipartite graph learning (EMOBGL) framework for robust and discriminative multiview clustering.
• To address limitations in capturing neighborhood structures, high-order correlations, and anchor-anchor interactions in existing MVC methods.

Main Methods:

• Constructs a second-order bipartite graph (SOBG) using first-order similarity (FOS) and second-order similarity (SOS) to capture local and neighboring relationships.
• Incorporates tensor Schatten-p regularizer for multiorder bipartite graph (MOBG) construction to capture third-order similarity (TOS) across views.
• Introduces anchor structure regularization (ASR) to model anchor-anchor interactions, enhancing graph expressiveness and discriminability.

Main Results:

• The EMOBGL model effectively integrates multiorder and multiview relationships, achieving robust and discriminative clustering performance.
• Experiments on 13 benchmark datasets show EMOBGL consistently outperforms 13 state-of-the-art methods in clustering accuracy and robustness.
• An efficient alternating direction method of multipliers (ADMMs) is developed for model optimization, with proven convergence to a KKT stationary point.

Conclusions:

• EMOBGL provides a unified framework for multiview clustering that surpasses existing methods by considering richer structural information.
• The proposed method demonstrates superior performance and robustness across various datasets, offering a significant advancement in MVC.
• The availability of source code facilitates further research and application of the EMOBGL framework.