An Urban Traffic Flow Fusion Network Based on a Causal Spatiotemporal Graph Convolution Network
Round 1
Reviewer 1 Report
The authors present convolution structure adopted by causal convolution gated linear unit (C-GLU) extracts the time characteristics of traffic flow. This is to avoid the problem of long running time caused by implemented recursive network. Thay have made a good review in the field of learning methods.
I consider the mathematical notation of the road traffic network model to be correct. The range of the calculation sample size and the results confirm the potential of the proposed method.
To sum up, I believe that the work is worth publishing in this journal.
Author Response
Dear reviewer,
It is a great honor to have your recognition of this work. Thank you for your
review and wish you have a good day!
Best regards
Author Response File: 
Author Response.pdf
Reviewer 2 Report
The paper presents trafic flow prediction based on spatio-temporal GCN. It is a current interesting topic in urban traffic. The paper has been presented well in general but it is lack in detail of methodology and results. The coments are as follow.
1.The figure in the introduction (figure 1) can be placed in the next section such in material and method. Figure 1 needs to be added the units of the vertical axis and the horizontal axis.
2. Lately, there are many publications related to traffic flow with spatio-temporal approach, the authors need to add discussion to that recently published by adding references for making comprehensive overview to the novelty of the paper.
Zhong, W., Suo, Q., Jia, X., Zhang, A. and Su, L., 2021, July. Heterogeneous Spatio-Temporal Graph Convolution Network for Traffic Forecasting with Missing Values. In 2021 IEEE 41st International Conference on Distributed Computing Systems (ICDCS) (pp. 707-717). IEEE.
Siddiqi, M.D., Jiang, B., Asadi, R. and Regan, A., 2021. Hyperparameter Tuning to Optimize Implementations of Denoising Autoencoders for Imputation of Missing Spatio-temporal Data. Procedia Computer Science, 184, pp.107-114.
Joelianto, E., Fathurrahman, M.F., Sutarto, H.Y., Semanjski, I., Putri, A. and Gautama, S., 2022. Analysis of Spatiotemporal Data Imputation Methods for Traffic Flow Data in Urban Networks. ISPRS International Journal of Geo-Information, 11(5), p.310.
and others
3. The defined graph of the road network need to be displayed in order to shown and the related feature.
4. What specific improvements should the authors consider regarding the methodology? What further controls should be considered?
5. How the method will overcome the temporal feature of different days?
6.It is not clear how the spatio-temporal of the method works
7. It is difficult for the reader to follow your paper without to consult the references. The authors need to make improvement.
8. Figure 8 showed the main feature of the proposed method but the high frequency characteristic of the traffic flow totally loss.
9. How to conclude the performance of the intersection network from the results related to traffic parameter metrics, e.g. congestion, travel time or others.
Author Response
Thank you for your comments.Please see the attachment.
Author Response File: Author Response.pdf
Reviewer 3 Report
The paper proposes a spatial-temporal graph neural network for the traffic prediction task. The topic is of practical interest and experiments show leading results. However, the motivation and equations which leads to the author’s “low-pass filter” graph convolutional neural are very clear. Especially the following claim is confusing:
“the high-order form of Chebyshev graph convolution is actually a highpass filter, and the high-order values will enhance the weight of high-frequency features.”
The reviewer suggests a better mathematical explanation of it. One thing in literature worth noting is that “About ChebNet, one can argue that the convolutions mostly cover the spectrum, none of the kernels focuses on some certain parts of the spectrum ” [1]
Also, the author may want to add an ablation study to demonstrate the improvement from the proposed LPF convolution and other network components.
Minor：
- There are some grammar issues and typos, e.g., in the related work, the beginning of section 3.1, etc.
- Fig. 4 is structured well. The authors may want to add indices such as Step 1, Step 2, etc., or simply put all steps horizontally in one row.
[1] M. Balcilar, G. Renton, P. Héroux, B. Gauzere, S. Adam, and P. Honeine, “Bridging the gap between spectral and spatial domains in graph neural networks,” arXiv preprint arXiv:2003.11702, 2020.
Author Response
Thank you for your comments.Please see the attachment.
Author Response File: Author Response.pdf
