Multi-view graph clustering via dual attention fusion and collaborative optimization.

Area of Science:

• Data Mining
• Machine Learning
• Graph Theory

Background:

• Multi-view graph clustering leverages multiple data sources for node partitioning.
• Existing methods struggle with unique structural information and cross-view relationships.
• Lack of semantic consistency leads to unstable representations and poor clustering quality.

Purpose of the Study:

• To propose a novel end-to-end method, Multi-view Graph Clustering via Dual attention fusion and Collaborative optimization (MGCDC), for improved multi-view graph clustering.
• To effectively capture unique structural information within each view and complementary relationships across views.
• To enforce global semantic consistency for stable consensus representations and enhanced clustering quality.

Main Methods:

• Encoding each view using a graph attention autoencoder for view-specific node embeddings.
• Integrating embeddings via view-level attention for a unified consensus representation.
• Employing cross-view cluster alignment and semantic consistency enhancement losses for collaborative optimization.

Main Results:

• MGCDC achieves highly competitive performance on five benchmark datasets.
• The method effectively integrates multi-view information and refines cluster assignments.
• Demonstrated superior clustering quality compared to state-of-the-art methods.

Conclusions:

• MGCDC offers a robust solution for multi-view graph clustering.
• The dual attention fusion and collaborative optimization strategy enhances clustering performance.
• The proposed method effectively addresses limitations of existing approaches.