Deep Graph Translation.

Area of Science:

• Graph-based network analysis
• Machine learning for network science
• Generative modeling for complex systems

Background:

• Deep generative models excel at graph modeling but lack control over output.
• Existing models cannot condition graph generation on a source graph.

Purpose of the Study:

• To propose a novel conditional generative model for graph translation.
• To enable controlled transformation of source graphs into target graphs.

Main Methods:

• Introduced Graph-Translation-Generative-Adversarial-Nets (GT-GAN).
• Developed a graph translator with novel convolution/deconvolution layers for global and local feature learning.
• Proposed a conditional graph discriminator for training.

Main Results:

• GT-GAN significantly outperforms baseline methods in effectiveness and scalability.
• Achieved at least 32.5% improvement in functional connectivity prediction for brain networks over SOTA methods.
• Demonstrated strong performance on cybernetworks, IoT, and neuroscience datasets.

Conclusions:

• GT-GAN offers a powerful approach for controlled graph generation and translation.
• The model advances network analysis in diverse scientific and engineering domains.
• GT-GAN provides a scalable and effective solution for conditional graph synthesis.