ACDm: Autoregressive-conditioned diffusion model for future state generation in temporal knowledge graph reasoning

Temporal Knowledge Graphs (TKGs) reasoning aims to forecast absent future events within Knowledge Graphs (KGs) that evolve over time, which is crucial for understanding the dynamic evolution of knowledge. However, the inherent complexity of the real world implies that relying solely on historical semantics to predict future states presents limitations. To address this challenge, we propose the Autoregressive-Conditioned Diffusion Model (ACDm), a novel approach for TKG reasoning. ACDm is designed with a bidirectional iterative diffusion process conditioned on historical information, which is modeled by autoregressive methods. The model simulates the distribution of entities and relations at future timestamp and generates representations that closely align with the actual future, thereby achieving superior reasoning performance. Our evaluations across several benchmark datasets (ICEWS14, ICEWS18, GDELT, and WIKI) demonstrate that ACDm achieves state-of-the-art performance, with MRR improvements of 1.8 % on ICEWS14 and 0.9 % on GDELT over strong baselines like RPC and TiRGN. This work underscores the importance of modeling uncertainty and dynamic evolution in TKGs, offering a flexible and robust framework for future event prediction under real-world variability.