Research map / Graph learning / Knowledge and temporal graphs
Clustering then Propagation: Select Better Anchors for Knowledge Graph Embedding
Read the paperCheck your idea against 108,967 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 RecPiece, an anchor-based knowledge graph embedding method that selects anchors by clustering triplet features into as many clusters as relation types and mapping representative triplets to anchor entities.
Problem
Shallow KGE models scale poorly in parameters and memory, and existing anchor-based KGE methods select anchors through primitive strategies (random, degree centrality, PageRank), limiting performance.
Approach
Cluster triplet features (not entity features) with k-means using the number of relation types as cluster count, pick triplets nearest each centroid in proportion to relation frequency, and map them to anchor entities that feed the NodePiece-style hashing and MLP propagation.
Key results
- On FB15k-237, RecPiece + RotatE reaches MRR 0.265 vs 0.254 for NodePiece + RotatE with 2.9M vs 3.2M parameters.
- On OGB WikiKG 2, RecPiece + AutoSF reaches MRR 0.598 with 5.9M parameters, versus 0.570 for NodePiece + AutoSF (6.9M) and 0.584 for PairRE+LRE (505M).
- On WD50K (5% labeled), RecPiece + CompGCN improves Hard Acc to 0.538 from 0.513 for NodePiece + CompGCN with 0.64M vs 0.75M parameters.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace NodePiece anchor selection with relational clustering-based selection | RotatE · FB15k-237 link prediction | MRR: 0.254 -> 0.265 | Helped |
| Replace NodePiece anchor selection with relational clustering-based selection | RotatE · WN18RR link prediction | MRR: 0.396 -> 0.402 | Helped |
| Replace NodePiece anchor selection with relational clustering-based selection | RotatE · CoDEx-L link prediction | MRR: 0.190 -> 0.198 | Helped |
| Replace NodePiece anchor selection with relational clustering-based selection | RotatE · YAGO 3-10 link prediction | MRR: 0.231 -> 0.243 | Helped |
| Replace NodePiece anchor selection with relational clustering-based selection | COMPGCN · WD50K (5% labeled) entity classification | PRC-AUC: 0.443 -> 0.459 | Helped |
| Add relational clustering-based anchor selection to AutoSF | AutoSF · OGB WikiKG 2 link prediction | MRR: 0.570 ±0.003 -> 0.598 ±0.003 | Helped |
| Use relational clustering-based anchor selection instead of random selection (ablation) | RecPiece · FB15k-237 entity prediction | MRR: 0.249 -> 0.265 | Helped |
| Use relational clustering-based anchor selection instead of node degree centrality (ablation) | RecPiece · FB15k-237 entity prediction | MRR: 0.250 -> 0.265 | Helped |
| Use relational clustering-based anchor selection instead of Personalized PageRank (ablation) | RecPiece · FB15k-237 relation prediction | MRR: 0.878 -> 0.884 | Helped |
| Use KG structural information rather than PLM textual features for feature preparation (ablation) | RecPiece + RotatE · FB15k-237 link prediction | MRR: 0.262 -> 0.265 | Helped |
| Cluster triplet features instead of entity features (ablation) | RecPiece + RotatE · FB15k-237 link prediction | MRR: 0.259 -> 0.265 | Helped |
| Use BisectingKMeans instead of KMeans for clustering (ablation) | RecPiece · FB15k-237 link prediction | MRR | No clear effect |
Surprising findings
- Pretraining features with a PLM (BERT) gave weaker results than structure-based pretraining (MRR 0.262 vs 0.265 with KG-self features), suggesting structure matters more for link prediction.
- Entity-level clustering with best cluster number (10) underperformed triplet-level clustering, supporting the relation-count-based design.
Limitations
- Feature preparation relies on an independent pretrained auxiliary model (NodePiece/GraIL), so anchor quality depends on how well it is trained.
- Performance still lags shallow KGE backbones (e.g., RotatE 0.338 vs 0.265 MRR on FB15k-237).
- Anchor-based pipeline adds pretraining time and memory cost.
Experimental setup
Datasets: FB15k-237 WN18RR CoDEx-L YAGO3-10 WD50K OGB WikiKG 2
Benchmarks: FB15k-237 WN18RR CoDEx-L YAGO3-10 WD50K OGB WikiKG 2
Metrics: MRR Hits@10 ROC-AUC PRC-AUC Hard Accuracy parameter count Effi. (MRR/#P) memory running time
Baselines: TransE DistMult ComplEx PairRE RotatE TripleRE AutoSF NodePiece + RotatE NodePiece + AutoSF NodePiece + CompGCN LRE + PairRE
Builds on: NodePiece RotatE AutoSF CompGCN
Closest papers by meaning
- Knowledge Graph Completion with Relation-Aware Anchor Enhancement (AAAI 2025 · similarity 0.89)
- KGDM: A Diffusion Model to Capture Multiple Relation Semantics for Knowledge Graph Embedding (AAAI 2024 · similarity 0.89)
- KG-FIT: Knowledge Graph Fine-Tuning Upon Open-World Knowledge (NeurIPS 2024 · similarity 0.88)
- Towards Continual Knowledge Graph Embedding via Incremental Distillation (AAAI 2024 · similarity 0.87)
- Select and Augment: Enhanced Dense Retrieval Knowledge Graph Augmentation (Abstract Reprint) (AAAI 2024 · similarity 0.87)
- Structure Is All You Need: Structural Representation Learning on Hyper-Relational Knowledge Graphs (ICML 2025 · similarity 0.87)
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.