摘要
为解决数据的时间特征和空间特征表征不完备、忽视未知节点推断补全和不确定性估计问题,提出了具有不确定性感知的图Transformer神经过程(graph Transformer neural process,Graphformer NP)时空数据补全方法。采用局部图卷积神经网络和Transformer学习空间与时间特征的联合确定性表示,引入神经过程,通过潜在状态转换学习缺失位置的潜在变量,补全缺失值并获得不确定性估计。充分的时空特征表征提高了补全的准确度,未知节点的推断补全弥补了传感器稀疏部署的不足,不确定性估计提高了在实际应用中决策部署的可靠性。在多个数据集上的实验验证了该方法的准确性和有效性,并为不确定性估计提供了可靠性参考。
To address the issues of incomplete representation of temporal and spatial features,neglect of inference completion for unknown nodes,and lack of uncertainty estimation,a spatio-temporal data imputation method with uncertainty-aware capabilities,named the graph Transformer neural process(GraphformerNP),is proposed.This method utilizes a local graph convolutional neural network(GCN)and Transformer to learn the joint deterministic representation of spatial and temporal features.By incorporating a neural process,it learns latent variables for missing locations through latent state transitions,enabling the completion of missing values and providing uncertainty estimation.Adequate spatiotemporal feature representation improves the accuracy of imputation,inference completion of unknown nodes compensates for the problem of sparse sensor deployment,and uncertainty estimation enhances the reliability of deployment decisions in real-world critical applications.The experiments on multiple datasets have verified the accuracy and effectiveness of the method,and provided valuable reliability references for uncertainty estimation.
作者
孙菲艳
郝文宁
曲爱妍
靳大尉
程恺
SUN Feiyan;HAO Wenning;QU Aiyan;JIN Dawei;CHENG Kai(College of Command&Control Engineering,Army Engineering University of PLA,Nanjing 210007,China;School of Software Engineering,Jinling Institute of Technology,Nanjing 211169,China)
出处
《陆军工程大学学报》
2025年第4期71-79,共9页
Journal of Army Engineering University of PLA
基金
国家自然科学基金(61806221)。
