Graph Neural Networks and Graph Transformers
Neural networks that learn from nodes, edges and their features for node classification, link prediction and whole-graph property prediction.
Tier Grades
Tier A: Frontier AI labs
feasible
Microsoft, Stanford and Mila publish graph architectures (GraphGPS, Graphormer).
Tier B: Indonesian AI labs
feasible
Linear-complexity models (GCN, GraphGPS) need modest compute.
Tier C: Institution units such as ITB
feasible
GCN came from a 2-author team, a scale any institution unit can match.
Tier D: Organisations such as SEMAI (PoT, PoC, PoV)
feasible
Fraud rings, supply networks and grids are graph problems a delivery team can prototype with GCN-class models; client graph data is the limit.
Where this comes from
Tiers: a) frontier AI labs, b) Indonesian AI labs, c) institution units such as ITB, d) organisations such as SEMAI delivering proof of technology, concept and value. Each rationale cites the anchors below.
Anchors
- ComputeGCN scales linearly in the number of graph edges; GraphGPS gives a graph transformer with complexity linear in nodes and edges, O(N+E).
- DataGCN: citation networks and a knowledge graph dataset; GraphGPS: a diverse set of graph benchmarks.
- Team size (proxy)Proxy only (author count of an exemplar paper, not a headcount): 2 authors on GCN; 6 on GraphGPS.
- Time horizonYears.
- Shelf-life signalsGraph transformers (GraphGPS, Graphormer) moving past message passing on benchmarks, Domain graph models such as GridSFM for power grids
- Pipeline roleTier A and universities publish architectures; Tier C applies them to Indonesian networks; Tier D prototypes fraud, logistics and grid use cases.
Where this comes from
Team size is a proxy (author counts of exemplar papers), not a headcount.
Who and What
Exemplars
Observed actors
Stanford University
Microsoft AI
University of Amsterdam
Mila
University of Cambridge
Technion - Israel Institute of Technology
Massachusetts Institute of Technology
Stony Brook UniversityLinked models
GraphGPSgraphexemplarGraphormergraphexemplarGraphSAGEgraphexemplarGCN (Graph Convolutional Network)graphexemplarGridSFM Opengraphproduced byGraphormer base (PCQM4Mv2)graphproduced byGATv2graphproduced by
TGN (Temporal Graph Networks)graphproduced by
HGT (Heterogeneous Graph Transformer)graphproduced by
LightGCNrecommendationused by
NGCF (Neural Graph Collaborative Filtering)recommendationused byGIN (Graph Isomorphism Network)graphproduced byGAT (Graph Attention Network)graphproduced byR-GCN (Relational Graph Convolutional Network)graphproduced bynode2vecgraphproduced byDeepWalkgraphproduced byLinks
R&D TopicGraph Neural Networks and Graph Transformers
ModelsGCN (Graph Convolutional Network)University of Amsterdam, graphGraphGPSMila, graphGraphSAGEStanford University, graphGraphormerMicrosoft AI, graph+12 more
GCN (Graph Convolutional Network)University of Amsterdam, graphGraphGPSMila, graphGraphSAGEStanford University, graphGraphormerMicrosoft AI, graph+12 moreUse Casesnone recorded
Who Runs Itnone recorded
Models 16
GAT (Graph Attention Network)University of Cambridge, Mila, graphproduced byNGCF (Neural Graph Collaborative Filtering)National University of Singapore, University of Science and Technology of China (USTC), recommendationused byGATv2Technion - Israel Institute of Technology, graphproduced byGIN (Graph Isomorphism Network)Massachusetts Institute of Technology, Stanford University, graphproduced byGraphGPSMila, Universite de Montreal, graphexemplarGraphormerMicrosoft AI, graphexemplarGraphormer base (PCQM4Mv2)Microsoft AI, graphproduced byHGT (Heterogeneous Graph Transformer)UCLA, Microsoft Research, graphproduced byLightGCNUniversity of Science and Technology of China (USTC), recommendationused byR-GCN (Relational Graph Convolutional Network)University of Amsterdam, graphproduced byDeepWalkStony Brook University, graphproduced byGCN (Graph Convolutional Network)University of Amsterdam, graphexemplarPapers
- Semi-Supervised Classification with Graph Convolutional NetworksThomas N. Kipf et al., ICLR 2017 (per arXiv comment), 2016 founding
- Recipe for a General, Powerful, Scalable Graph TransformerLadislav Rampášek et al., NeurIPS 2022 (per arXiv comment), 2022 state of the art
- Do Transformers Really Perform Bad for Graph Representation?Chengxuan Ying et al., NeurIPS 2021, 2021 state of the art
- How Powerful are Graph Neural Networks?Keyulu Xu et al., arXiv preprint, 2018 state of the art
- A Comprehensive Survey on Graph Neural NetworksZonghan Wu et al., arXiv preprint, 2019 survey
Where this comes from
The papers to read first: where the direction began, its current best, and a survey when one exists.
Sources
- arxiv.org/abs/1609.02907 https://arxiv.org/abs/1609.02907
- arxiv.org/abs/2205.12454 https://arxiv.org/abs/2205.12454
To cite this page: Project AURION, AI Use-Cases Research & Innovation, "Graph Neural Networks and Graph Transformers", https://aurion.semai.id/topics/graph-neural-networks-and-graph-transformers/, records of October 2026.