Research map / Graph learning / Knowledge and temporal graphs
KGDM: A Diffusion Model to Capture Multiple Relation Semantics 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 KGDM, a conditional DDPM that recasts knowledge graph entity prediction as conditional entity generation in embedding space, using an MLP-based denoiser, to capture multiple relation semantics.
Problem
Knowledge graph embedding methods make deterministic entity predictions and fixed embeddings, so they fail to capture the multiple semantics of relations (one-to-many, many-to-many), degrading knowledge graph completion; existing probabilistic embeddings are computationally heavy.
Approach
Casts entity prediction as conditional generation: a DDPM diffuses learnable entity/relation embeddings, and a Conditional Entity Denoiser (MLP blocks with a TransE/RotatE-style scoring module on head and relation embeddings) generates tail entity embeddings, trained with a negative-sampling-style loss and ranked by L1 distance.
Key results
- On FB15k-237, KGDM reaches MRR 0.520 (H@1 42.3), versus 0.385 for the best embedding baseline AnKGE-HAKE, a 43.6% relative gain.
- On Kinship, KGDM reaches MRR 0.789 and on UMLS MRR 0.909, improving the best non-embedding methods by 6.1 and 6.7 points respectively.
- Gains are largest on 1-N and N-N relation types on FB15k-237 (e.g., N-N MRR 0.465 vs 0.275 for CompGCN in tail-pred table), while WN18RR improvement is marginal (0.516 vs 0.500).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use KGDM instead of NBFNet | KG completion · FB15k-237 | MRR: 0.415 -> 0.520 | Helped |
| use KGDM instead of NBFNet | KG completion · WN18RR | MRR: 0.551 -> 0.516 | Hurt |
| use KGDM instead of CIBLE | KG completion · Kinship | MRR: 0.728 -> 0.789 | Helped |
| use KGDM instead of RNNLogic | KG completion · UMLS | MRR: 0.842 -> 0.909 | Helped |
| add the Scoring Module (ablation) | KGDM · FB15k-237 | MRR: 0.449 -> 0.520 | Helped |
| add the Scoring Module (ablation) | KGDM · Kinship | MRR: 0.635 -> 0.789 | Helped |
| replace two transformer layers with MLP layers in the CEDenoiser (ablation) | KGDM · FB15k-237 | MRR: 0.334 -> 0.520 | Helped |
| replace two transformer layers with MLP layers in the CEDenoiser (ablation) | KGDM · Kinship | MRR: 0.661 -> 0.789 | Helped |
Surprising findings
- Replacing the MLP blocks in the denoiser with transformer layers drops FB15k-237 MRR from 0.520 to 0.334, so the simpler MLP is preferred for short triplets.
- Gains on WN18RR are marginal, which the authors attribute to its high entity-to-relation ratio making target distribution harder to model.
Limitations
- Performance improvement on WN18RR is not significant due to high entity-to-relation ratio.
- Relies on choosing different scoring functions per dataset (addition for FB15k-237, Hadamard for WN18RR/Kinship/UMLS) based on pattern types.
Experimental setup
Datasets: FB15k-237 WN18RR Kinship UMLS
Benchmarks: FB15k-237 WN18RR Kinship UMLS
Metrics: MRR Hits@1 Hits@3 Hits@10 MR
Baselines: TransE RotatE TransH TransG KG2E KG2GM ComplEx-N3 TuckER ConvE HAKE DualE SEA
Builds on: DDPM TransE RotatE negative sampling loss
Closest papers by meaning
- KG-FIT: Knowledge Graph Fine-Tuning Upon Open-World Knowledge (NeurIPS 2024 · similarity 0.90)
- Clustering then Propagation: Select Better Anchors for Knowledge Graph Embedding (NeurIPS 2024 · similarity 0.89)
- Towards Global-Topology Relation Graph for Inductive Knowledge Graph Completion (AAAI 2025 · similarity 0.89)
- KGGen: Extracting Knowledge Graphs from Plain Text with Language Models (NeurIPS 2025 · similarity 0.88)
- Knowledge Graph Completion with Relation-Aware Anchor Enhancement (AAAI 2025 · similarity 0.87)
- Towards Continual Knowledge Graph Embedding via Incremental Distillation (AAAI 2024 · 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.