Multigraph Fusion for Dynamic Graph Convolutional Network.

Area of Science:

• Machine Learning
• Graph Neural Networks
• Data Representation

Background:

• Graph convolutional networks (GCNs) excel at representation learning by leveraging data structure.
• GCN robustness is limited by feature matrix quality and initial graph integrity.
• Existing methods struggle with optimal graph construction for GCNs.

Purpose of the Study:

• To propose a novel multigraph fusion method for improving GCN robustness.
• To generate a high-quality graph and a low-dimensional data space for GCNs.
• To enhance classification performance through improved graph and feature representation.

Main Methods:

• A multigraph fusion technique is proposed to integrate common and complementary information from multiple local graphs.
• A unified local graph is created and fused with a global graph to form the initial graph for GCN.
• The method employs a dual graph fusion process for simultaneous low-dimensional space learning and intrinsic graph structure discovery.

Main Results:

• Experimental validation on real datasets demonstrated the efficacy of the proposed method.
• The approach significantly outperformed existing comparison methods in classification tasks.
• The multigraph fusion method successfully produced high-quality graphs and effective low-dimensional representations.

Conclusions:

• The proposed multigraph fusion method enhances GCN robustness and performance.
• The dual graph fusion framework effectively learns both data representation and graph structure.
• This approach offers a promising direction for improving GCNs in various applications.