Research map / Graph learning / Knowledge and temporal graphs
Structure Is All You Need: Structural Representation Learning on Hyper-Relational Knowledge Graphs
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 MAYPL, a purely structure-based message passing model for hyper-relational knowledge graphs that learns structure-driven initial embeddings and attentive fact-level message passing, supporting inductive inference on new entities and relations.
Problem
Link prediction (reasoning) on hyper-relational knowledge graphs, where existing methods underuse graph structure and cannot handle new entities or relations at inference time.
Approach
Learns to initialize entity and relation representations from the HKG's co-occurrence and position structure, then applies attentive neural message passing over facts (fact-level messages aggregated to entities and relations with position-specific matrices), using no textual or external features.
Key results
- On WikiPeople- (predicting all entities), MAYPL reaches MRR 0.521 vs. 0.509 for HAHE in the transductive setting, and on WD50K (all entities) MRR 0.411 vs. 0.402 for HAHE.
- On inductive HKG benchmark WD20K(100)v1, MAYPL achieves MRR 0.486 vs. 0.113 for StarE; on WD20K(100)v2 MRR 0.298 vs. 0.067 for the best QBLP variant.
- On the 12 KG inductive datasets of Lee et al. (2023b), MAYPL has total loss 0.000 vs. 2.300 for InKGE-EIPR (sum of total losses 0.572 vs. 3.328).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use MAYPL's structure-driven initializer rather than learnable vectors for each entity and relation (ablation) | MAYPL · WikiPeople´ | MRR: 0.388 -> 0.521 | Helped |
| Add attentive neural message passing (ablation) | MAYPL · WikiPeople´ | MRR: 0.109 -> 0.521 | Helped |
| Use attentive aggregation instead of mean aggregation (ablation) | MAYPL · WikiPeople´ | MRR: 0.504 -> 0.521 | Helped |
| Aggregate relation representations from incident entity-fact pairs rather than use a simple projection (ablation) | MAYPL · WikiPeople´ | MRR: 0.519 -> 0.521 | Helped |
| Use Hadamard products instead of concatenation in message computation (ablation) | MAYPL · WikiPeople´ | MRR: 0.518 -> 0.521 | Helped |
| Share parameters for query and entity representations (ablation) | MAYPL · WikiPeople´ | MRR: 0.520 -> 0.521 | Helped |
| Use the structure-driven initializer rather than learnable vectors for each entity and relation (ablation) | MAYPL · WK-50 | MRR: 0.061 -> 0.109 | Helped |
| Add attentive neural message passing (ablation) | MAYPL · WK-50 | MRR: 0.029 -> 0.109 | Helped |
| Use attentive aggregation instead of mean aggregation (ablation) | MAYPL · WK-50 | MRR: 0.086 -> 0.109 | Helped |
| Aggregate relation representations from incident entity-fact pairs rather than use a simple projection (ablation) | MAYPL · WK-50 | MRR: 0.062 -> 0.109 | Helped |
| Use Hadamard products instead of concatenation in message computation (ablation) | MAYPL · WK-50 | MRR: 0.090 -> 0.109 | Helped |
| Share parameters for query and entity representations (ablation) | MAYPL · WK-50 | MRR: 0.081 -> 0.109 | Helped |
Surprising findings
- A purely structure-based model without textual or external features outperforms text-augmented baselines such as QBLP on inductive HKG tasks.
- Hyper-relational facts cannot be converted losslessly into n-ary relational representations or knowledge hypergraphs, so the authors argue for direct HKG handling.
Limitations
- Attention computation (Bahdanau-style) demands substantial computing resources and limits scalability.
- Training time on WikiPeople- (20h) is longer than some baselines such as Hy-Transformer (5h).
Experimental setup
Datasets: WD50K WikiPeople- WikiPeople NL-50 WK-50 FB-50 WD20K(100)v1 WD20K(100)v2 WP-IND MFB-IND
Benchmarks: WD50K WikiPeople- WikiPeople NL-50 WK-50 FB-50 WD20K(100)v1 WD20K(100)v2 WP-IND MFB-IND
Metrics: MRR Hit@10 Hit@1 Hit@3
Baselines: StarE GRAN Hy-Transformer HyperFormer HAHE ShrinkE HyNT InGram HyRel InKGE-EIPR HCNet QBLP
Builds on: StarE HAHE NBFNet InGram
Closest papers by meaning
- Knowledge Graphs Can be Learned with Just Intersection Features (ICML 2024 · similarity 0.87)
- Knowledge Graph Completion with Relation-Aware Anchor Enhancement (AAAI 2025 · similarity 0.87)
- Towards Global-Topology Relation Graph for Inductive Knowledge Graph Completion (AAAI 2025 · similarity 0.87)
- Clustering then Propagation: Select Better Anchors for Knowledge Graph Embedding (NeurIPS 2024 · similarity 0.87)
- Text2NKG: Fine-Grained N-ary Relation Extraction for N-ary relational Knowledge Graph Construction (NeurIPS 2024 · similarity 0.87)
- Replacing Paths with Connection-Biased Attention for Knowledge Graph Completion (AAAI 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.