A Multi-condition diffusion model for completing hyper-relational knowledge graphs via conditional entity generation

Incomplete hyper-relational knowledge graphs: Hyper-relational knowledge graphs remain incomplete and need additional prediction tasks for enhancement.

Deterministic prediction limits semantic diversity: Existing deterministic completion methods restrict the ability to model diverse semantics from varying entity and relation roles.

Ambiguous target identification with many candidates: Identifying prediction targets from single semantic results is difficult and degrades performance when many candidate answers exist.

Qualifier-aware diffusion conditioning: Direct diffusion models struggle with arbitrary numbers of qualifiers and with leveraging multiple conditions to capture latent semantics.

Incomplete hyper-relational knowledge graphs: Hyper-relational knowledge graphs remain incomplete and need additional prediction tasks for enhancement.

Conditional entity generation: The paper reformulates entity prediction as a conditional entity generation problem.

Multi-Conditional Diffusion Model: The paper proposes MCDM, a diffusion-based method for conditional entity generation in hyper-relational knowledge graph completion.

Denoising diffusion target distribution estimation: MCDM uses a denoising diffusion process to estimate the probability distribution of target entities.

Multi-condition denoising module: The method introduces a multi-condition denoising module with dataset-specific conditional aggregation functions to generate targets and manage qualifiers.

Conditional entity generation: The paper reformulates entity prediction as a conditional entity generation problem.

State-of-the-art outperformance:

Up to 10 percent improvement:

Probabilistic generation framing: