Graph Clustering With Graph Capsule Network.

Area of Science:

• Data Science
• Machine Learning
• Network Analysis

Background:

• Graph clustering is crucial for partitioning graphs with similar structures.
• Deep learning methods have advanced graph clustering but face limitations.
• Existing methods struggle with effective graph property representation and isolated embedding/clustering processes.

Purpose of the Study:

• To propose a novel capsule-based graph clustering (CGC) algorithm.
• To overcome limitations of current deep graph clustering methods.
• To learn cluster-oriented graph embeddings effectively.

Main Methods:

• Constructed a graph clustering capsule network (GCCN) using capsules to capture graph properties.
• Designed an iterative optimization strategy to update GCCN parameters and clustering assignments.
• Developed a method to integrate embedding learning and clustering processes.

Main Results:

• The proposed CGC algorithm demonstrated superior performance over existing methods.
• Achieved high scores in standard evaluation metrics: ACC, NMI, and ARI.
• Visualization confirmed the effectiveness of capsules in learning cluster-oriented embeddings.

Conclusions:

• The capsule-based graph clustering algorithm offers a significant improvement in graph clustering tasks.
• The GCCN effectively captures graph properties and learns cluster-oriented embeddings.
• The integrated approach enhances the suitability of learned embeddings for clustering.