摘要
针对现有时序知识图谱推理模型不能充分挖掘时序知识图谱中并发事实的结构依赖和潜在关系,时间编码方式单一不合理,获取时间信息单调的问题。提出了一种基于联合卷积的时序知识图谱推理模型。该模型一方面通过引入使用联合聚合器的图卷积神经网络挖掘节点邻域信息的表面语义和潜在特征;另一方面,通过对时间进行向量编码和事件属性编码来捕获丰富的时间信息,增强模型的时间敏感性。在ICEWS14、ICEWS05-15、YAGO和GDELT数据集上的实验结果表明,模型在MRR、Hits@1、Hits@3和Hits@10上普遍优于基线模型,同时关系预测也均优于基线模型。
Existing temporal knowledge graph reasoning models fail to fully explore the structural dependencies and potential relationships among concurrent facts in temporal knowledge graphs.Additionally,these models often rely on simplistic and unreasonable time encoding methods,resulting in inadequate temporal information acquisition.This paper proposes a temporal knowledge graph reasoning model based on joint convolution.This model uses a joint aggregator in a graph convolutional neural network to capture the surface semantics and latent features of node neighborhood information.It also employs vector and event attribute encoding for time to capture rich temporal information,enhancing the temporal sensitivity of the model.Experimental results on ICEWS14,ICEWS05-15,YAGO,and GDELT datasets demonstrate that the model consistently outperforms baseline models in MRR,Hits@1,Hits@3,and Hits@10,as well as in relation prediction.
作者
张成珅
马汉达
ZHANG Chengshen;MA Handa(School of Computer Science and Communication Engineering,Jiangsu University,Zhenjiang,Jiangsu 212013,China)
出处
《计算机工程与应用》
北大核心
2025年第18期166-174,共9页
Computer Engineering and Applications
基金
镇江市重点研发计划项目(GY2023034)。
关键词
时序知识图谱
图卷积神经网络
门控循环单元
联合卷积
temporal knowledge graph
graph convolutional network
gated recurrent unit
joint convolution
