Reproduce “GraphRNN Revisited: An Ablation Study and Extensions for Directed Acyclic Graphs” on OpenResearch

About this paper

Taniya Das, Mark Koch, Maya Ravichandran, Nikhil Khatri

Researchers from the Universities of Oxford and Cambridge reproduced the GraphRNN model, achieving MMD scores comparable to or better than originally reported for graph generation. An ablation study clarified that Breadth-First Search (BFS) ordering is crucial for performance, and M-truncation is advantageous when it provides substantial prediction space reduction, while an extension using topological sort for Directed Acyclic Graphs (DAGs) outperformed a generic directed graph approach.

Key insights

• The Breadth-First Search (BFS) ordering mechanism is a fundamental component of GraphRNN, essential for regularizing graph representations and maintaining high generative performance.
• The `M`-truncation feature is most effective and provides clear benefits when it substantially reduces the dimensionality of the prediction space, offering less impact when the reduction is minimal.
• Integrating domain-specific structural properties, such as acyclicity via topological sorting, yields superior results for generating constrained graph types like DAGs compared to more generic approaches.

Summary by alphaXiv.