An arXiv version 1 preprint dated Aug. 25, 2026, reports an AI method that assigns newly arriving knowledge-graph entities to semantic clusters from raw features, without retraining or graph access at assignment time. On MAG, its projector completed assignment in milliseconds, while BMGC-Retrain and DEMM-Retrain required more than 65 seconds per batch; the reported cluster quality was close to the retraining reference.
SNAP-KG is trained offline with view-specific graph neural networks and a Transformer encoder that fuses relation-specific embeddings through contrastive learning. At inference, the projector maps an incoming entity to an embedding and assigns it to the nearest cluster center. That step requires neither retraining nor graph reconstruction for each arrival.
The test bed
Tests covered five benchmark graph datasets: ACM with 3,025 nodes, DBLP with 4,057, IMDB with 4,780, YELP with 2,614 and MAG with 113,919. A separate OGB-WikiKG2 evaluation began with 2.5 million entities and 535 relation types, used the eight highest-frequency relation types and retained 2.4 million nodes after preprocessing.
For ACM, DBLP and YELP, streaming tests held out 200 nodes in two batches of 100. For MAG, 2% of nodes, or 2,278 samples, were withheld in five batches.
Speed and clustering
On MAG, SNAP-KG reported the best accuracy and F1 score among the listed methods, at 68.20% and 68.35%. Its reported accuracy was 91.50% on ACM, 91.82% on DBLP and 91.20% on YELP, within 2% of the strongest transductive baselines; on IMDB, NMI and ARI were low across methods.
The scores also differed sharply between the full multi-view model and the cited single-view variants: NMI was 71.08% versus 41.28% on ACM, and 74.03% versus 5.02% on DBLP.
The search trade-off
SNAP-KG uses the assigned cluster to narrow the candidates considered for entity resolution, or matching duplicate records, and link prediction, or scoring possible links, rather than searching the full set. The paper reports 62% to 75% fewer candidates across the five benchmark datasets and 97% fewer on OGB-WikiKG2.
In a cross-dataset test, a model trained on DBLP and applied to arriving ACM entities reduced candidates by 76.2%. Entity-resolution F1 was 92.64%, compared with 93.45% for a full scan, a difference of 0.81 percentage points.
Results varied by task. On ACM, cluster-scoped entity-resolution precision, recall and F1 were all 100%, matching the full scan, with a reported 65.4% candidate reduction. On DBLP, cluster-scoped link-prediction AUPR was 83.01% versus 75.75% for a full scan; on MAG, cluster-scoped AUC-ROC was 92.21% versus 96.43%, and AUPR was 37.95% versus 61%.
With K set to 250 clusters on OGB-WikiKG2, candidate reduction was 97.75%. Entity-resolution F1 was 80.67%, while link-prediction AUC-ROC was 98.24% and AUPR was 31.13%, compared with 13.35% for the full-scan baseline. That comparator sampled 500,000 nodes rather than scanning the complete graph.
The boundary problem
The main warning concerns links that cross cluster boundaries. In the MAG scoping analysis, 14,938 of 75,578 ground-truth positive edges fell outside the scoped candidates; 19.8% of true edges crossed cluster boundaries, and 45 of 2,254 arriving nodes, or 2.0%, had no positive edge left.
The comparisons are reported as point estimates, with no confidence intervals or significance tests. Entity-resolution queries were mainly synthetic, and the MAG analysis shows why cluster scoping can be vulnerable to losing links that cross boundaries.
Paper data and sources
Original title: SNAP-KG: Streaming Node Assignment via Projection for Knowledge Graph Entity Integration
Authors: Jui-Chien Lin, Mohammad Mohammadi Amiri, Oshani Seneviratne
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-25
DOI: Not available
Original paper · Full text