A Simple and Scalable Representation for Graph Generation

Overview

We address limitations in neural network-based graph generation by introducing GEEL (gap encoded edge list), a compact representation whose size corresponds to the number of edges rather than being quadratic in the node count. Our method incorporates gap encoding and bandwidth restriction to minimize vocabulary size, supports autoregressive generation with node positional encoding, and extends to attributed graphs through custom grammar. Evaluation across twelve graph generation tasks demonstrates improved scalability and performance.

BibTeX

@inproceedings{
  jang2024a,
  title={A Simple and Scalable Representation for Graph Generation},
  author={Yunhui Jang and Seul Lee and Sungsoo Ahn},
  booktitle={The Twelfth International Conference on Learning Representations},
  year={2024},
  url={https://openreview.net/forum?id=nO344avRib}
}