时空融合的时序知识图谱多跳推理模型OA

To address the issues of semantic disconnection caused by separated entity-relation embedding spaces and limited temporal expressiveness in existing temporal knowledge graph multi-hop reasoning models,this paper proposed a spatiotempo-ral fusion multi-hop reasoning model(SF-MR).The model incorporated a triple distributor with dual-path residual connections and spatial convolution to capture cross-space semantic dependencies between entities and relations.It introduced a spatiotem-poral attention mechanism to jointly model entity temporal evolution and spatial correlations,dynamically fused via a gated net-work.A hierarchical reinforcement learning framework decoupled reasoning into relation-level and entity-level decisions,alle-viating action space explosion.Experiments on four benchmark datasets(ICEWS14,ICEWS18,WIKI,YAGO)demonstrate that SF-MR outperforms state-of-the-art baselines across multiple metrics.Specifically,on ICEWS14,SF-MR improves MRR,hits@3,and hits@10 by 1.10％,1.53％,and 2.69％,respectively,over the best baseline.Consistent improvements of 0.79％to 1.01％are observed on WIKI and YAGO.Ablation studies confirm the effectiveness of the triple distributor and spatiotemporal attention in enhancing semantic interaction and temporal modeling.