Browse Teach Curate

Token-Sequence Autoregressive Graph Generation with Transformers

A paradigm for generating graphs by serializing each graph into a sequence of discrete tokens that encode its nodes and edges, then generating new graphs one token at a time with an autoregressive Transformer trained by next-token prediction, as in language modeling. This sequence-based encoding is more compact than adjacency-matrix or diffusion-based approaches and yields a pretrained foundation model that can be fine-tuned for goal-oriented (property-targeted) generation and graph property prediction.

Rosie Text

Graph Generative Pre-trained Transformer Xiaohui Chen 1 Yinkai Wang 1 Jiaxing He 2 Yuanqi Du 3 Soha

The Graph Generative Pre-trained Transformer (G2PT) generates graphs by representing each graph as a sequence of tokens that encode its nodes and edges, then producing that sequence one token at a ti…