TSA-HGNN: a stability-aware multi-scale temporal graph neural network for dynamic community detection.

Area of Science:

• Network Science
• Graph Theory
• Machine Learning

Background:

• Dynamic community detection is crucial for understanding evolving networks.
• Existing methods often fail due to static graph assumptions and unstable community assignments.
• Real-world networks exhibit dynamic changes in nodes and edges, necessitating adaptive detection methods.

Purpose of the Study:

• To propose a stability-aware, multi-scale temporal graph neural network (TSA-HGNN) for dynamic community detection.
• To enhance the stability and accuracy of community detection in evolving graphs.
• To address the limitations of static graph algorithms in dynamic network environments.

Main Methods:

• Utilized GraphSAGE for snapshot-level spatial embedding.
• Employed a Temporal Convolutional Network (TCN) for short-term patterns and an Informer for long-term patterns.
• Incorporated an Echo State Network (ESN) as an auxiliary component and added a temporal smoothness constraint.

Main Results:

• TSA-HGNN achieved superior performance on benchmark datasets, outperforming state-of-the-art algorithms.
• Reported accuracy values of 0.9843, 0.9755, and 0.9931 on key datasets.
• Demonstrated high F-Score (0.9807), Modularity Q (0.8273), NMI (0.8878), and ARI (0.8677), indicating robust and stable results.

Conclusions:

• TSA-HGNN provides an efficient and highly stable solution for dynamic community detection.
• The framework effectively models multi-scale temporal patterns in evolving graphs.
• Experimental validation confirms the superiority and stability of TSA-HGNN over existing methods.