融合关系路径和上下文的归纳关系预测模型OA北大核心

The existing inductive relation prediction methods primarily focus on the relation paths between the different entities,often neglecting the properties of head and tail entities in relation context.To address this,we propose an inductive relation prediction model fusing relation paths and context(IRP-RPC),which incorporates the relation context as a complement to the relation path for inductive relation prediction.This model relies solely on relation semantics,enabling it to naturally generalize to fully inductive settings.First,we use a random walk-based pathfinding strategy to obtain relation paths and relation contexts.Then,we design and implement a hierarchical Transformer architecture,enhanced with a gated network,to unify and aggregate the relation paths and contexts,capturing both the relations between entities and the intrinsic attributes of entities.An adaptive weighted combination of these components is used to make the final prediction.Experiments on eight versions of inductive datasets from FB15K-237 and NELL-995 demonstrate that IRP-RPC model outperforms nine baseline models in terms of area under the precision-recall curve(AUC-PR)and hits@10 metrics,validating its effectiveness and scalability.The experimental results show that by fusing relation paths and relation contexts,the IRP-RPC model can more comprehensively capture the semantic relationships and structural information between entities,offering a significant advantage over traditional inductive methods,which suffer from under-utilization of path and context information.