摘要
随着大数据分析和深度学习的迅猛发展,时间序列预测方法被广泛应用于医学、金融、气象和交通等领域,为众多应用任务提供决策支持.针对小样本医学数据特征维度低和现有深度学习方法易于造成过拟合问题,研究小样本医学时间序列预测任务,提出融合趋势嵌入和粒度增强网络的预测方法.首先在卷积神经网络的框架下,粒度增强网络分别从时间维度和特征维度将医学时间序列数据提升为三维张量;然后以医学时间序列样本的一阶差分作为方向向量,基于方向导数生成趋势嵌入表征;再构建静态空间邻接矩阵和动态时间邻接矩阵,并通过时空图卷积网络学习时空嵌入表征;最后将构建的时空嵌入、预测嵌入和趋势嵌入整合到基于图卷积网络、门控循环单元和残差网络的网络架构之中,实现医学时间序列预测.在Cancer,ILI,Baries和COVID-19这4个数据集上的实验结果表明,与预测结果最佳的基线模型T-GCN相比,所提方法在每个数据集的MAE,MAPE和RMSE这3个评价指标上分别降低34.0607,0.0107,70.6728;11.1808,0.0950,20.7285;0.3546,0.1127,0.4553和449.2437,0.0144,1174.7273,其性能优于基线方法,验证了该方法的可行性及有效性.
With the rapid development of big data analysis and deep learning,time series forecasting methods have been widely used in domains of medicine,finance,meteorology,transportation,and so on,which provide various types of decision supports for many application tasks.To solve the problem of low feature dimension ubiquitously in medical data and the problem of overfitting occurred by small samples in the existing deep learning frameworks,this paper focuses on the task of medical time series forecasting with small samples,and develops a new framework with trend embedding and granularity enhancement network.First,a granularity enhancement module is constructed with tricks of convolution operations,which is used to augment the dimensions of the samples of medical time series respectively in time-steps and data features for generating three-dimensional tensor.Second,the first-order difference of each sample is taken as the direction vector to learn the corresponding trend embedding by mapping with its direction derivative.Third,a static spatial adjacency matrix and a dynamic temporal adjacency matrix are simultaneously constructed to learn the spatiotemporal embedding by a spatiotemporal Graph Convolutional Network(GCN).Finally,the constructed modules of spatial-temporal embedding,prediction embedding and trend embedding are integrated together into the architectural framework of the GCN,the gated recurrent unit and the residual network,which is trained to achieve the goal of medical time series forecasting.Extensive experimental results on four datasets including Cancer,ILI,Baries and COVID-19 show that the proposed approach in this paper reduces the three evaluation metrics(MAE,MAPE,RMSE)by 34.0607,0.0107,70.6728;11.1808,0.0950,20.7285;0.3546,0.1127,0.4553;449.2437,0.0144,1174.7273,compared with the best baseline model T-GCN.This illustrates the performance of the proposed method is superior than those of the baseline methods,indicating its feasibility and validation.
作者
于敬楠
张春霞
薛新月
薛晓军
牛振东
Yu Jingnan;Zhang Chunxia;Xue Xinyue;Xue Xiaojun;Niu Zhendong(School of Computer Science and Technology,Beijing Institute of Technology,Beijing 100081)
出处
《计算机辅助设计与图形学学报》
EI
CSCD
北大核心
2024年第6期948-959,共12页
Journal of Computer-Aided Design & Computer Graphics
基金
国家自然科学基金(62072039)。
关键词
时间序列预测
趋势嵌入
粒度增强网络
时空图卷积网络
time series forecasting
trend embedding
granularity enhancement network
spatiotemporal graph convolutional network
