Learning Graph Embedding With Adversarial Training Methods.

Area of Science:

• Graph Neural Networks
• Machine Learning
• Data Mining

Background:

• Graph embedding methods often prioritize structure or reconstruction, neglecting latent code distribution.
• Suboptimal embedding distributions can lead to inferior representations for graph analytics.

Purpose of the Study:

• To propose a novel adversarially regularized framework for graph embedding.
• To enhance graph representation learning by enforcing latent code distribution matching.

Main Methods:

• Utilized graph convolutional networks (GCNs) as encoders to embed topological and content information.
• Implemented adversarial training to align latent codes with prior Gaussian or uniform distributions.
• Developed two variants: Adversarially Regularized Graph Autoencoder (ARGA) and its variational version (ARVGA).

Main Results:

• The proposed ARGA and ARVGA models effectively learn graph embeddings.
• Experimental results demonstrated superior performance compared to 12 link prediction and 20 graph clustering algorithms.

Conclusions:

• Adversarial regularization is a powerful technique for improving graph embedding quality.
• The developed framework offers a promising approach for advanced graph analytics tasks.