基于图小波卷积神经网络的时空图挖掘模型

Spatio-temporal Graph Mining Model Based on Graph Wavelet Convolutional Neural Network

针对传统时空图网络模型对时空序列数据空间结构刻画和时空特性挖掘不充分的问题,提出一种基于图小波神经网络的时空图挖掘模型(ST-GWNN)。基于图小波神经网络通过学习节点特征的局部化表达来捕捉时空序列数据中的空间拓扑结构,时间门控卷积层通过门控线性单元所堆叠的因果卷积来提取时间特征信息,并将多个时间步的空间图相融合来学习时间和空间2个维度关联特征的能力,以更好地捕获时空序列中复杂的时空相关性信息。在公共交通数据集PEMS-BAY上的实验结果表明,ST-GWNN模型能够获得较好的预测效果,当预测时长为15 min时,在MAE、RMSE、MAPE 3个评价指标上相较于基准模型取得最小值,且较基准模型最优值分别降低了2.31%、6.96%、5.84%;当预测时长为30 min和60 min时,较基准模型最优的MAPE、RMSE值分别降低了4.9%、3.51%和6.05%、6.68%,可适用于图网络属性的时空关系预测任务。

To address the insufficiency of traditional spatio-temporal graph network models in describing the spatial structure and mining of the spatio-temporal characteristics of spatio-temporal series data,this study proposes a Spatio-Temporal graph mining model based on a Graph Wavelet convolutional Neural Network(ST-GWNN).The model adopts a graph wavelet neural network to capture the spatial topology in spatio-temporal sequence data by learning the localized representation of node features.The temporally gated convolutional layer extracts temporal feature information by stacking causal convolutions of gated linear units.In addition,the model fuses the spatial graphs of multiple time steps and can simultaneously learn the correlation features of the two dimensions of time and space,thus better enabling complex spatio-temporal correlation information to be captured in the spatio-temporal sequence.Experimental results on the public transport dataset PMS-Bay show that the ST-GWNN model has better prediction effects.When the prediction time is 15 min,the ST-GWNN model can obtain the minimum values of Mean Absolute Error(MAE),Root Mean Square Error(RMSE),and Mean Absolute Percentage Error(MAPE),which are 2.31%,6.96%,and 5.84%less,respectively,than the optimal values of the benchmark model.When the prediction time is 30 and 60 min,the MAPE and RMSE decrease by 4.9%and 3.51%and by 6.05%and 6.68%,respectively,as compared with those of the benchmark model,which can be applied to the spatio-temporal relationship prediction task of graph network attributes.