Auto-regressive Graph Generation Modeling with Improved Evaluation Methods

Graphs are frequently used to organize and represent information in various realworld domains such as social sciences, medicine, and language. Unfortunately, generating realistic graphs and evaluating them quantitatively is often notoriously difficult due to the complex structural dependencies inherent to graphs. In this paper, we make two central contributions: first, we propose a novel classifier-based evaluation method to measure graph generation quality and diversity. Secondly, we demonstrate empirically that our attention-based Transformer models achieve competitive performance in terms of generation quality compared to state of the art recurrent models.