Multigranularity Deep Graph Convolutional Neural Network Node Clustering Leveraging Spatial Information.

Area of Science:

• Artificial Intelligence
• Data Science
• Machine Learning

Background:

• Clustering graph-structured and empty graph-structured data is crucial for data analysis.
• Existing methods struggle with data lacking explicit topology and often ignore spatial information.
• Graph convolutional neural networks (GCNs) show promise but face challenges with topology-deficient data.

Purpose of the Study:

• To develop a generalizable clustering method for both graph and empty graph-structured data.
• To address the limitations of shallow networks and the neglect of spatial information in current methods.
• To enhance the interpretability and quality of similarity matrices in clustering.

Main Methods:

• Proposed a multigranularity deep GCN node clustering method (CMDGCN) leveraging spatial information.
• Converted empty graph data to graph data using the k-nearest neighbor (k-nn) algorithm.
• Constructed multigranularity graph structures and improved the self-expressiveness principle for enhanced similarity matrices.

Main Results:

• CMDGCN effectively handles both graph-structured and empty graph-structured data.
• The method demonstrated superior performance across multiple datasets compared to existing techniques.
• Achieved high-quality and interpretable similarity matrices by incorporating original graph structure.

Conclusions:

• The proposed CMDGCN offers a robust and effective solution for graph node clustering and empty graph-structured data analysis.
• This work provides novel perspectives and tools for handling diverse graph data types.
• The method's effectiveness and robustness were validated through extensive experiments.