Dual generative adversarial graph networks: Unsupervised and semi-supervised learning with spherical graph...

Area of Science:

• Graph-based data analysis
• Machine learning
• Network science

Background:

• Graph-structured data is vital for complex systems.
• Attributed graph clustering and semi-supervised node classification are key challenges.
• Existing methods may struggle with noise and robustness.

Purpose of the Study:

• To propose a novel dual generative adversarial framework for attributed graph clustering.
• To extend this framework for semi-supervised node classification.
• To enhance the robustness and performance of graph embeddings.

Main Methods:

• Utilizing a dual generative adversarial network (GAN) structure.
• Incorporating a probabilistic encoder with a von Mises-Fisher mixture model (vMFMM) for spherical graph embeddings.
• Extending the clustering framework with a mixture density network for semi-supervised learning.

Main Results:

• The proposed framework achieves stable and superior performance in attributed graph clustering.
• Spherical graph embeddings offer explicit control over representation trade-offs.
• The dual GAN structure effectively removes latent space noise, improving robustness.
• The semi-supervised extension enhances node classification accuracy.

Conclusions:

• The dual generative adversarial framework provides a principled and effective approach to graph clustering and node classification.
• The method demonstrates significant advantages over established baselines.
• The integration of vMFMM and dual GANs offers a robust solution for graph data analysis.