Researchers propose DMKGC, a framework that applies conditional diffusion models to transfer knowledge across multiple domains for knowledge graph completion. The method treats each knowledge graph as a partial view of entity information, generating domain-general embeddings conditioned on support graphs while preserving domain-specific details.
- Initializes domain-agnostic entity embeddings as priors and encodes them within individual KGs.
- Fuses equivalent entities from support KGs to guide conditional diffusion generation.
- Uses prior embeddings as proxy objectives to ensure unbiased generation across domains.
- Achieves a 4.3% average MRR improvement in tail entity prediction over state-of-the-art methods on 14 KGs in 3 benchmarks.
The approach addresses the risk of suppressing domain-specific contextual information inherent in consistency-constrained methods, demonstrating sustained gains particularly in low-resource data settings.