Covariance-driven consistency and MMD complementarity with hybrid graph for multi-view clustering

Area of Science:

• Machine Learning
• Data Science
• Computer Vision

Background:

• Effective multi-view clustering requires learning unified representations that capture both shared structures and unique information across different data views.
• Existing methods often struggle to balance consistency and complementarity, limiting the quality of the integrated representation.

Purpose of the Study:

• To propose a novel multi-stage framework for multi-view clustering that learns a high-quality unified representation.
• To effectively capture consistent structural relationships and complementary information across multiple views.

Main Methods:

• Constructing K-nearest neighbours (KNN) graphs and hypergraphs to capture local geometric structures at multiple scales.
• Employing an Edge-Boost fusion strategy for a robust unified affinity graph, followed by Graph Neural Network (GNN) for feature refinement.
• Integrating a complementarity loss using Maximum Mean Discrepancy (MMD) and a Covariance-driven Basis Projection (CBP) module for adaptive aggregation.

Main Results:

• The proposed framework successfully learns a unified representation that balances consistent and complementary information.
• The Edge-Boost fusion and CBP module enhance the robustness and discriminative power of the final embedding.
• The integration of MMD loss encourages views to learn differentiated information, improving overall clustering performance.

Conclusions:

• The multi-stage framework provides an effective approach to multi-view clustering by learning a superior unified representation.
• The proposed methods for graph fusion, feature refinement, and adaptive aggregation offer significant improvements over existing techniques.
• This work contributes to advancing multi-view clustering by addressing the challenge of integrating diverse information sources effectively.