SCGG: A deep structure-conditioned graph generative model

Area of Science:

• Artificial Intelligence
• Machine Learning
• Graph Theory

Background:

• Deep learning excels at graph data modeling for real-world problems.
• Conditional generation enhances graph modeling by meeting specific criteria.
• Graph completion is a challenging problem in recovering missing graph components.

Purpose of the Study:

• To present a conditional deep graph generation method, SCGG.
• To enable generation of graph samples based on structural conditions.
• To address the problem of graph completion.

Main Methods:

• SCGG utilizes an initial subgraph for autoregressive generation of nodes and edges.
• The architecture combines a graph representation learning network and an autoregressive generative model.
• The graph representation network captures long-range node dependencies for structural conditioning.

Main Results:

• SCGG effectively generates new graph samples conditioned on substructures.
• The model successfully addresses graph completion by recovering missing nodes and edges.
• The method demonstrates linear computational complexity with respect to the number of graph nodes.

Conclusions:

• SCGG offers a novel approach to conditional deep graph generation.
• The method outperforms state-of-the-art baselines on synthetic and real-world datasets.
• SCGG provides an efficient solution for graph completion and related tasks.