【论文阅读】A Hybrid Model Integrating Local and Global...Traffic Prediction[一种融合局部和全局空间相关性的交通预测混合模型]（4）

原文地址：https://ieeexplore.ieee.org/abstract/document/9667362

4. Experiments（实验）

A. Experiments Settings（实验设置）

1) Datasets（数据集）

We evaluated the performance of T-LGGCN through two real highway datasets: the PEMS04 dataset and the PEMS08 dataset. Both datasets are from the Caltrans Performance Evaluation System. We took the traffic speed data as the traffic flow information. PEMS04 was collected from San Francisco Bay. It contains 307 sensors, and the time is from January 1, 2018, to February 28, 2018. PEMS08 was from San Bernardino. It contains 170 sensors, and the time is from July 1, 2016, to August 31, 2016. The traffic data in both datasets were aggregated every 5 minutes. There were 288 records per day and no missing data in both datasets. We used the threshold Gaussian kernel to calculate the adjacency matrix of both datasets. During the experiments, we use 80% of the data as the training set and 20% for testing. The normalization operation is performed first when the data is input.
我们通过两个真实的高速公路数据集:PEMS04数据集和PEMS08数据集评估了T-LGGCN的性能。这两个数据集都来自Caltrans性能评估系统。我们将交通速度数据作为交通流信息。PEMS04产自旧金山湾。它包含307个传感器，时间从2018年1月1日到2018年2月28日。PEMS08产自圣贝纳迪诺。它包含170个传感器，时间为2016年7月1日至2016年8月31日。两个数据集的流量数据每5分钟聚合一次。每天有288条记录，两个数据集都没有丢失数据。我们使用阈值高斯核计算两个数据集的邻接矩阵。在实验过程中，我们使用80%的数据作为训练集，20%用于测试。规范化操作首先在数据输入时执行。

2) Parameter Settings（参数设置）

The experiments were implemented based on the TensorFlow framework. We used the Adam optimizer to optimize the model. Adam is an optimization algorithm designed to find the global optimal points. It combines the advantages of Momentum and RMSProp to minimize the loss function [48]. Adam Optimizer is an optimizer that implements Adam algorithm in TensorFlow which has been validated on a large number of neural network experiments. After repeated experiments, when the model performance reached the optimum, the main parameters were set as follows: $\alpha$ is set to $0.8$, the GRU dimension is set to $64$, the input step is set to $12$, the learning rate is set to $0.001$, the batch size is set to $32$, and the training epoch is set to $100$.
实验基于TensorFlow框架实现。我们用 Adam优化器 来优化模型。Adam是一个优化算法，用来寻找全局最优点。它结合了Momentum和RMSProp的优点，最小化了 [48] 的损失函数。Adam 优化器是一个在TensorFlow中实现Adam算法的优化器，该算法已被大量神经网络实验验证。经过反复实验,当模型的性能达到最优,主要参数设置如下: $\alpha$ 设置为 $0.8$ , GRU维度设置为 $64$ , 输入设置为12, learning rate（学习率）设置为 $0.001$ , batch（训练批量）大小设置为 $32$ , training epoch（训练轮数）被设置为 $100$ 。

3) Evalution Metrics（评估矩阵）

We use the root mean square error (RMSE), mean absolute error (MAE), accuracy, coefficient of determination ($R^2$) as the evaluation metrics, and the detailed definitions are as follows.
我们使用均方根误差 (RMSE)、平均绝对误差 (MAE)、accuracy (准确率)、决定系数 ($R^2$)作为评价指标，详细定义如下。
$$\begin{array}{rlr}RMSE& =\sqrt{\frac{1}{IS}\sum _{i=1}^I\sum _{s=1}^S(y_i^s-{y_i^s}^{\prime }{)}^2}& (15)\\ MAE& =\frac{1}{IS}\sum _{I=1}^I\sum _{s=1}^S|y_i^s-{y_i^s}^{\prime }|& (16)\\ Accuracy& =1-\frac{{\Vert Y-Y^{\prime }\Vert }_F}{{\Vert Y\Vert }_F}& (17)\\ R^2& =1-\frac{{\sum }_{i=1}^I{\sum }_{s=1}^S(y_i^s-{y_i^s}^{\prime }{)}^2}{{\sum }_{i=1}^I{\sum }_{s=1}^S(y_i^s-\bar{y}{)}^2}& (18)\end{array}$$
where $I$ represents the input time steps, $S$ is the numbers of sensors, $y_i^s$ and ${y_i^s}^{\prime }$ denote the real traffic speed data and predicted data respectively, $Y$ and $Y^{\prime }$ are the sets of $y_i^s$ and ${y_i^s}^{\prime }$ , $\bar{y}$ is the mean value of $Y$.
其中，$I$ 表示输入的时间步长，$S$ 为传感器个数，$y_i^s$ 和 ${y_i^s}^{\prime }$ 分别表示实际交通速度数据和预测数据， $Y$ 和 $Y^{\prime }$ 分别为 $y_i^s$ 和 ${y_i^s}^{\prime }$ 的集合，$\bar{y}$ 为 $Y$ 的平均值。

B. Baselines（基线）

To demonstrate the effectiveness of our model, we select parametric methods (i.e., HA and ARIMA), non-parametric methods (i.e., SVR), and deep learning methods (i.e., GCN, GRU, and T-GCN) for comparison.
为了证明我们模型的有效性，我们选择了参数方法(即 HA 和 ARIMA )、非参数方法(即 SVR )和深度学习方法(即 GCN、GRU 和 T-GCN)进行比较。

 (1) HA [15]: HA is the simplest parametric method. The average value of historical traffic parameters is used as the model prediction result in our experiments.
 (1) HA [15]: HA是最简单的参数方法。实验中采用历史交通参数的平均值作为模型预测结果。

 (2) ARIMA [16]: The basic idea of ARIMA is to treat the time-series data as a random series and describe this series approximately with a certain mathematical model. A smoothness test is performed on the data first, and an ARMA model is fitted for prediction. We set the value of the autoregressive coefficient to 1, and the values of the difference order and the moving average to 0.
 (2) ARIMA [16]: ARIMA的基本思想是将时间序列数据视为随机序列，用一定的数学模型对该序列进行近似描述。首先对数据进行平滑检验，然后拟合ARMA模型进行预测。我们设置自回归系数为1，差序和移动平均的值为0。

 (3) SVR [18]: SVR is one of the common nonparametric prediction methods. The data are mapped to a multidimensional space using a nonlinear function, and then linear regression is performed on them.
 (3) SVR [18]: SVR是常用的非参数预测方法之一。利用非线性函数将数据映射到多维空间，然后对其进行线性回归。

 (4) GCN [34]: GCN is a neural network method for spatial correlation analysis. A one-layer graph convolutional network is used, and the specific calculation process is detailed in Equation (4).
 (4) GCN [34]: GCN是一种用于空间相关性分析的神经网络方法。采用单层图卷积网络，具体计算过程如式 (4) 所示。

 (5) GRU [6]: GRU is a variant of RNN. It is usually used to analyze the time-series data. The data is directly input into GRU, and the calculation process is shown in 3.4. We set the number of hidden layer neurons to 64.
 (5) GRU [6]: GRU是RNN的变体。它通常用于时间序列数据的分析。数据直接输入到GRU中，计算过程如3.4所示。我们设置隐藏层神经元的数量为64个。

 (6) Temporal Graph Convolutional Network(T-GCN) [33]: T-GCN is a hybrid neural network. It consists of one-layer GCN and GRU. The feature data is input into GCN for calculation, and the output of the GCN layer is input into GRU to get the prediction resualts.
 (6)时间图卷积网络(Temporal Graph Convolutional Network, T-GCN) [33]: T-GCN是一种混合神经网络。它由单层GCN和GRU组成。将特征数据输入到GCN进行计算，将GCN层的输出输入到GRU中得到预测结果。

C. Results of Traffic Prediction

Table 1 shows the prediction performance of our model and other baseline models on the PEMS04 and PEMS08 datasets for 5 minutes prediction task. It can be seen that T-LGGCN Compared with the basic model T-GCN, the RMSE error of our model is reduced by $12.1\%$ on PEMS04 and by $23.1\%$ on PEMS08.
表1显示了我们的模型和其他基线模型在 PEMS04 和 PEMS08 数据集上5分钟预测任务的预测性能。可以看出，与基本模型 T-GCN 相比，我们的模型在 PEMS04 和 PEMS08 上的RMSE误差分别降低了 $12.1\%$ 和 $23.1\%$ 。

TABLE 1. Prediction results of different models on PEMS04 and PEMS08.
表1. 不同模型对PEMS04和PEMS08的预测结果

 From the overall prediction effect, the neural network methods have the best performance, and the non-parametric method SVR is weaker than the neural network but better than the parametric methods ARIMA and HA. It is shown that the neural network methods can better learn the non-smoothness of traffic flow. It can deeply explore internal features of the traffic flow by training a large amount of data and effectively improve the prediction accuracy.
 从整体预测效果来看，神经网络方法的预测效果最好，非参数方法 SVR 比神经网络弱，但优于参数方法 ARIMA 和 HA。结果表明，神经网络方法能更好地学习交通流的非平滑性。通过训练大量数据，可以深入挖掘交通流内部特征，有效提高预测精度。

 The RMSE of GRU on the PEMS04 and PEMS08 is $13.8\%$ and $32.5\%$ lower than the RMSE of GCN, respectively. GRU is used to capture temporal information. It can be found that temporal correlation is one of the important features of traffic flow. When we handle traffic prediction task, we need to take the temporal correlation into consideration.
 在PEMS04和PEMS08上，GRU的RMSE分别比GCN的RMSE低 $13.8\%$ 和 $32.5\%$ 。GRU用于捕获时间信息。可以发现，时间相关性是交通流的重要特征之一。在处理交通预测任务时，我们需要考虑时间相关性。

 We can also see that the hybrid neural networks outperform the single neural networks. Compared with GCN and GRU, the RMSE of T-LGGCN on the PEMS04 is improved by $36.2\%$ and $27.4\%$. This indicates that the hybrid model takes into account both temporal and spatial features, which can more comprehensively explore the traffic flow characteristics.
 我们还可以看到，混合神经网络优于单一神经网络。与GCN和GRU相比，T-LGGCN在PEMS04上的RMSE分别提高了 $36.2\%$ 和 $27.4\%$ 。这说明混合模型兼顾了时间和空间特征，能够更全面地探索交通流特征。

 It is noted from the table that T-LGGCN has a good ability to mine global and local spatial features. Compared with the T-GCN, the RMSE of T-LGGCN decreased by $12.1\%$ and $23.1\%$ on the PEMS04 and PEMS08. T-GCN only aggregates the spatial information of adjacent first-order sensors. T-LGGCN combines the spatial features of the node itself, first-order neighboring sensor spatial information, and long- range highly correlated sensor spatial information so as to analyze the spatial correlation comprehensively.
 从表中可以看出，T-LGGCN具有良好的全局和局部空间特征挖掘能力。与T-GCN相比，T-LGGCN在PEMS04和PEMS08上的RMSE分别降低了 $12.1\%$ 和 $23.1\%$ 。T-GCN只聚合相邻一阶传感器的空间信息。T-LGGCN结合节点本身的空间特征、一阶相邻传感器空间信息和远程高度相关的传感器空间信息，综合分析空间相关性。

D. Prediction Performance of T-LGGCN

To have a deeper understanding of the T-LGGCN performance, we visualize the prediction results of one sensor data on two test sets separately, as shown in Fig. 5. We visualize the prediction results for one day and one week on two datasets. It is observed that our model fits better.
为了更深入地了解T-LGGCN性能，我们将一个传感器数据在两个测试集上的预测结果分别可视化，如图5所示。我们在两个数据集上可视化一天和一周的预测结果。据观察，我们的模型更适合。


FIGURE 5. The visualization results in test set. (a) T-LGGCN predictions for one day on PEMS04. (b) T-LGGCN predictions for one day on PEMS08. © T-LGGCN predictions for one week on PEMS04. (d) T-LGGCN predictions for one week on PEMS08.
图5. 测试集中的可视化结果。(a) T-LGGCN对PEMS04一天的预测。(b) T-LGGCN对PEMS08一天的预测。© T-LGGCN对PEMS04上一周的预测。(d) T-LGGCN对PEMS08一周的预测。

Meanwhile, we compare the prediction results of T-GCN with T-LGGCN, as shown in Fig. 6. Obviously, the fit of T-GCN is poorer than that of T-LGGCN. Although the T-GCN prediction results are mostly close to the true values, we can find that the T-GCN does not accurately capture the traffic flow characteristics when the traffic flow state changes drastically. The T-GCN fit curve behaves relatively smoothly in adjacent times (shown in the orange box). This is because T-GCN uses GCN to perform spatial feature aggregation operations, which may cause an over-smoothing problem, resulting in smoother predicted values at the peak. We improved the GCN spatial aggregation method, and it can be seen that our model effectively avoids this drawback.
同时，我们将T-GCN和T-LGGCN的预测结果进行对比，如图6所示。显然，T-GCN的拟合不如T-LGGCN。虽然T-GCN预测结果大多接近真实值，但我们可以发现，当交通流状态发生剧烈变化时，T-GCN并不能准确捕捉交通流特征。T-GCN拟合曲线在相邻时间内表现相对平稳(如橙色框中所示)。这是因为T-GCN使用GCN进行空间特征聚合操作，可能会造成过平滑问题，导致峰值预测值更加平滑。我们改进了GCN空间聚集方法，可以看到我们的模型有效地避免了这一缺陷。


FIGURE 6. The comparison of results in test set with T-GCN. (a) The comparison of results for one day on PEMS04. (b) The comparison of results for one day on PEMS08.
图6. 测试集结果与T-GCN的比较。(a) PEMS04一天的结果比较。(b) PEMS08一天的结果比较。

参考文献

[6] S. Bai, J. Z. Kolter, and V. Koltun, ‘‘An empirical evaluation of generic convolutional and recurrent networks for sequence modeling,’’ 2018, arXiv:1803.01271.

[15] X. Yin, G. Wu, J. Wei, Y. Shen, H. Qi, and B. Yin, ‘‘Deep learning on traffic prediction: Methods, analysis and future directions,’’ IEEE Trans. Intell. Transp. Syst., early access, Feb. 10, 2021, doi: 10.1109/ TITS.2021.3054840.

[16] Z. Zhang, Y. Li, H. Song, and H. Dong, ‘‘Multiple dynamic graph based traffic speed prediction method,’’ Neurocomputing, vol. 461, pp. 109–117, Oct. 2021, doi: 10.1016/j.neucom.2021.07.052.

[18] Q. Li, Z. Han, and X. M. Wu, ‘‘Deeper insights into graph convolutional networks for semi-supervised learning,’’ in Proc. 32nd AAAI Conf. Artif. Intell. (AAAI), 2018, pp. 3538–3545.

[33] X. Chen, H. Chen, Y. Yang, H. Wu, W. Zhang, J. Zhao, and Y. Xiong, ‘‘Traffic flow prediction by an ensemble framework with data denoising and deep learning model,’’ Phys. A, Stat. Mech. Appl., vol. 565, Mar. 2021, Art. no. 125574, doi: 10.1016/j.physa.2020.125574.

[34] L. Yuting, Z. Ming, M. Chicheng, B. Bo, Z. Zhiheng, Y. Kai, W. Guanghui, and Y. Guiying, ‘‘Graph neural network,’’ SCIENTIA SINICA Math., vol. 50, no. 3, p. 367, Mar. 2020, doi: 10.1360/n012019-00133.

[48] D. P. Kingma and J. L. Ba, ‘‘Adam: A method for stochastic optimization,’’ in Proc. 3rd Int. Conf. Learn. Represent. (ICLR) Conf. Track, 2015, pp. 1–15.