DDGC: A diffusion-based approach for dynamic graph clustering.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Graph Theory

Background:

• Static graph clustering methods struggle with evolving network structures and dynamic node categories.
• Existing dynamic graph clustering approaches often fail to address minority classes and emerging unknown classes simultaneously.

Purpose of the Study:

• To propose a novel diffusion model-based dynamic graph clustering method for evolving graphs with unknown classes.
• To enhance the identification of new and minority classes in dynamic graph clustering scenarios.

Main Methods:

• Integration of graph convolutional networks (GCN) with diffusion models.
• Leveraging kernel density estimation and Tweedie's formula for density regularization in embedding space.
• Utilizing data density fluctuations for unsupervised pseudo-label assignment.

Main Results:

• The proposed method outperforms state-of-the-art baselines on dynamic graph clustering tasks.
• Demonstrated superior performance in scenarios with emerging unknown classes and minority samples.
• Achieved adaptive class discovery and sample augmentation in both static and evolving graph environments.

Conclusions:

• The novel diffusion model effectively addresses limitations of existing methods in dynamic graph clustering.
• The framework provides a robust solution for online graph understanding and real-world dynamic applications.
• This work bridges the gap between static and dynamic graph clustering research.