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

Abstract Hyper-relational knowledge graphs incorporate qualifiers to assign contextual roles to entities and relations, but they remain incomplete and require additional prediction tasks for further enhancement. However, existing methods depend on deterministic prediction for completion, restricting their ability to capture the diverse semantics arising from the varying roles of entities and relations. Accurately identifying targets from single semantic results during prediction is challenging, leading to performance degradation, particularly with numerous candidate answers. To address this challenge, we reformulate the entity prediction task as a conditional entity generation problem and propose a novel method, the Multi-Conditional Diffusion Model (MCDM). Specifically, MCDM utilizes a denoising diffusion process to estimate the probability distribution of targets. Directly employing diffusion models fails to handle arbitrary numbers of qualifiers, and effectively leveraging these multiple conditions to capture diverse latent semantics remains a significant challenge. The multi-condition denoising module is introduced to help generate targets and efficiently manage qualifiers with dataset-specific conditional aggregation functions. Conceptually, our work advances hyper-relational knowledge graph completion from deterministic prediction to probabilistic conditional generation with diffusion models, and provides a principled way to exploit qualifier information for modeling diverse latent semantics. Extensive experiments on real-world datasets demonstrate that our approach significantly outperforms state-of-the-art baselines, achieving up to a 10% improvement over the runner-up methods.