The article ‘Towards Principled Graph Transformers’ discusses the expressive power of graph learning architectures, particularly the Edge Transformer, within the Weisfeiler Leman (k-WL) hierarchy. The study highlights the Edge Transformer’s ability to deliver strong predictive performance on real-world tasks, surpassing other theoretically aligned architectures. It also theorizes that the Edge Transformer has at least 3-WL expressive power. The authors propose a concrete implementation of the Edge Transformer for various graph learning tasks.
Publication date: 18 Jan 2024
Project Page: https://arxiv.org/abs/2401.10119
Paper: https://arxiv.org/pdf/2401.10119
