AI Paper Atlas

Research map / Graph learning / GNN architectures

Graph Inductive Biases in Transformers without Message Passing

Liheng Ma, Chen Lin, Derek Lim, Adriana Romero-Soriano, Puneet K. Dokania, Mark Coates, Philip Torr, Ser-Nam Lim

ICML 2023 · Graph learning › GNN architectures · approach: Graph rewiring and scalable message passing · 15 citations (OpenAlex)

Read the paper

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
add degree scaler (ablation)GRIT · ZINCMAE: 0.076 -> 0.059Improved
add learned update of RRWP in attention (ablation)GRIT · ZINCMAE: 0.066 -> 0.059Improved
use global attention instead of sparse attention (ablation)GRIT · ZINCMAE: 0.066 -> 0.059Improved
use degree scaler instead of Graphormer degree encoding (ablation)GRIT · ZINCMAE: 0.072 -> 0.059Improved
use GRIT attention instead of Graphormer attention (ablation)GRIT · ZINCMAE: 0.117 -> 0.059Improved
use RRWP instead of RWSE (ablation)GRIT · ZINCMAE: 0.081 -> 0.059Improved
use RRWP instead of SPDPE (ablation)GRIT · ZINCMAE: 0.067 -> 0.059Improved
increase RRWP order K from 2 to 21 (ablation)GRIT · ZINCMAE: 0.147 -> 0.059Improved
vary RRWP order K among 7-42 (ablation)GRIT · ZINCMAE: 0.063 (K=7) vs 0.059 (K=21)No clear effect
GRIT attention with RRWP vs other attention modules on k-hop attention learningsingle-layer attention · K-hop synthetic 1-hopMAE: .043 -> .001Improved
GRIT vs GPS on small datasetsGraph Transformer · CIFAR10accuracy: 72.298 -> 76.468Improved

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy