基于图自编码-生成对抗网络的路网数据修复

完整的交通路网数据是实现智能交通系统的前提，故本文提出一种基于图自编码-生成对抗网络的方法对路网中缺失数据进行修复。首先，通过降噪图变分自编码器提取路网缺失数据的时空特征，使其能最大程度捕获原始路网信息；其次，基于该时空特征利用生成对抗网络生成路网数据，加入重建损失并优化生成对抗网络的目标函数，实现对缺失数据的有效插补；最后，采用西雅图(Seattle)和加州(PEMS04)路网速度数据集，针对不同缺失类型和缺失率下的数据修复进行对比实验。当随机缺失率在 10% ~70%时，Seattle 数据集的 MAE 指标在 2.38~3.25 之间，PEMS04 数据集的 MAE 指标在 1.46~2.38 之间；当聚集缺失率在 10%~70%时，Seattle 数据集的MAE指标在2.51~2.82之间，PEMS04数据集的MAE指标在1.52~1.54之间。对比结果表明，本文提出的路网数据修复方法均优于BP、DSAE、BGCP等模型。

Road Network Data Repair Based on Graph Autoencoder-generative Adversarial Network

 The completeness of the traffic and road network data affects the operation of the intelligent transportationsystems. This paper proposes a method based on graph autoencoder-generative adversarial network to repair themissing data in the road network. First, the spatiotemporal features of the missing road network data are extractedthrough the denoising graph variational autoencoder which captures the original road network information to thegreatest extent. Then, based on the spatial-temporal features, the study uses the generative adversarial network togenerate repaired road network data, adds reconstruction Loss, and optimizes the objective function of the generation ofthe confrontation network. The effective interpolation of missing data is then realized. This study uses the Seattle(Seattle) and California (PEMS04) road network speed datasets to conduct comparative experiments on data restorationwith different missing types and missing rates. When the random missing rate is between 10% and 70% , the meanabsolute error(MAE) index of the Seattle dataset is between 2.38 and 3.25. The MAE index of the PEMS04 data set isbetween 1.46 and 2.38. When the aggregated missing rate is between 10% and 70%, the MAE index of the Seattle dataset is between 2.51 and 2.82. The MAE index of the PEMS04 data set is between 1.52 and 1.54. The comparison resultsshow that the proposed road network data restoration methods perform better than the backpropagation network(BP),denoising stacked auto-encoder(DSAE), bayesian gaussian Candecomp/Parafac(BGCP) and other models involved inthe comparison analy