AI Paper Atlas

Research map / Graph learning / GNN architectures

gHAWK: Structural Encoding for Scalable Training of Graph Neural Networks on Knowledge Graphs

Humera Sabir, Fatima Farooq, Ashraf Aboulnaga

NeurIPS 2026 · Graph learning › GNN architectures · approach: Simple and scalable GNNs

Read the paperCheck your idea against 137,217 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes gHAWK, a structural-encoding approach for scalable GNN training on knowledge graphs that avoids reliance on iterative message passing to learn graph structure.

Problem

Existing message-passing GNNs struggle to scale to large knowledge graphs because learning graph structure through iterative message passing is inefficient, especially under mini-batch training where nodes see only partial neighborhoods.

Approach

Encodes graph structure explicitly through structural encodings rather than relying on iterative message passing to learn it, targeting efficient mini-batch training on large KGs.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
add Bloom filter and TransE structural features on top of default text featuressix GNN backbones · OGB-MAG and MAG240Mtest accuracy: OGB-MAG +9.5 (ClusterGCN) to +17.2 (GAT); MAG240M +5.0 (GraphSAGE) to +16.8 (HGT)Improved
add Bloom filter alone (no TransE) to default features (ablation)ClusterGCN · MAG240Mtest accuracy: 61.13 -> 57.37Hurt
add Bloom filter alone (no TransE) to default features (ablation)R-GCN, GraphSAGE, GraphSAINT, GAT, HGT · OGB-MAGtest accuracy: GraphSAINT 46.81 -> 50.26; R-GCN 37.67 -> 44.51Improved
add TransE on top of Bloom filter (ablation)GraphSAINT · OGB-MAGtest accuracy: 50.26 -> 57.37Improved
replace default text features with random vectors plus Bloom and TransE (ablation)six GNN backbones · OGB-MAGtest accuracy: GraphSAINT 46.81 (Word2Vec only) -> 54.55 (structure only)Improved
replace RoBERTa features with structure-only gHAWK features (ablation)six GNN backbones · MAG240Mtest accuracy: R-GCN 68.00 (RoBERTa) vs 59.23 (structure only)Mixed
add gHAWK features to a message-passing-free MLPMLP · OGB-MAGtest accuracy: 31.95 -> 35.61Improved
add gHAWK features to SGC, SIGN and SGFormerSGC, SIGN, SGFormer · OGB-MAGtest accuracy: SGFormer 51.35 -> 56.46; SGC 38.42 -> 40.84; SIGN 37.54 -> 40.46Improved
add gHAWK features to GraphSAGEGraphSAGE · DBLP, ACM, IMDb (HGB)Micro/Macro-F1: DBLP Micro-F1 82.50 -> 97.52; ACM Macro-F1 91.20 -> 93.16; IMDb Micro-F1 65.28 -> 69.14Improved
add gHAWK structural features to RotatE decoder (no message passing)RotatE · OGB-WikiKG2test MRR: 0.4341 -> 0.6941Improved
add gHAWK features to message-passing backbonesR-GCN, GraphSAGE, CompGCN · OGB-WikiKG2test MRR: R-GCN 0.3851 -> 0.5957; GraphSAGE 0.5689 -> 0.6129; CompGCN 0.6256 -> 0.6562Improved
add gHAWK features to RotatE, by head-entity training frequencyRotatE · OGB-WikiKG2 head queries (single run)head MRR: 1-5 triples 0.0109 -> 0.3497; 1000+ triples 0.8989 -> 0.7380Mixed

Earlier methods this paper compares with

Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.

Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.

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