STN-GCN: Spatial and Temporal Normalization Graph Convolutional Neural Networks for Traffic Flow Forecasting
Round 1
Reviewer 1 Report
The reviewed paper presents an STN-GCN modelling framework for spatial-temporal normalized graphical convolutional neural networks related to urban traffic forecasting.
This is relevant to the functioning of modern metropolises. Indeed, improving traffic flow, especially urban public transport, improves the living conditions and satisfaction of the inhabitants of large cities.
In doing so, it should be noted that we observe a complex spatial-temporal correlation of traffic data, which makes traffic volume forecasting very difficult. Therefore, in order to assess this traffic volume, the authors described the theoretical basis of STN-GCN modelling for spatial-temporal normalized graphical convolutional neural networks. They carried out calculations using these networks and the results were compared with other methods.
There is a certain dissatisfaction with the conclusions presented by the authors. I believe that they should be significantly expanded. This would improve the scientific quality of the reviewed article.
The article should be published in the ELECTRONICS journal with the corrections made.
Author Response
Point 1: There is a certain dissatisfaction with the conclusions presented by the authors. I believe that they should be significantly expanded. This would improve the scientific quality of the reviewed article.
Response 1: Given the recommendations provided for my conclusions, I have made necessary adjustments accordingly. The following is the updated conclusion section.
In conclusion, this paper proposes a temporal normalized graph convolutional neural network (STN-GCN) model for traffic flow prediction. The model incorporates spatial-temporal normalization to preprocess the input data before passing it into the temporal convolution layer, facilitating the extraction of temporal features from urban road traffic data, particularly for medium and long-term prediction. Moreover, the Transformer architecture dynamically captures spatial dependencies across multiple scales, which further enhances the extraction of spatial features. Additionally, the integration of Curriculum Learning techniques improves the experimental results. In terms of prediction performance at different time intervals, the proposed model exhibits smoother transitions and consistently outperforms other models. Furthermore, it surpasses the benchmark model in terms of its ability to predict traffic flow over medium and long durations. As part of our future work, we plan to explore the integration of STN-GCN with other deep learning models to uncover latent structured features within the input data.
Please refer to the attachment for the revised manuscript.
Author Response File: 
Author Response.docx
Reviewer 2 Report
The manuscript “STN-GCN: Spatial and Temporal Normalization Graph Convolutional Neural Networks for Traffic Flow Forecasting” provides an interesting study of traffic flow prediction using a novel CNN model proposed by the authors. Although the study is not very novel, it demonstrates increased traffic flow prediction accuracy in comparison with other methods. The manuscript is well written, although I would recommend to describe in more details which parts of the graphs corresponds to roads, crossroads etc. More detailed discussion of temporal prediction capabilities would be also fruitful. I believe that several issues should be addressed before publication of the paper.
1) Although the authors give very brief explanation of traffic graph construction in section 3, more detailed explanation should be added. Please clarify what are the roads in the graph G, doe they correspond to nodes A? What do you mean by saying that Q is a speed and is takes as 1? Is it constant for all graph elements? How many past points should be evaluated in the equation (1) in order to obtain future signal Xt+T?
2) Using of low- and high- frequencies and decomposition of spatial characteristic onto “local” and “global” are a great idea, however the authors should discuss in more details which temporal and spatial limits on real world should be considered as local / global and low- / high- frequency.
3) The authors describe validation of the proposed model on two datasets, consisting of 4-month information of traffic flow. The authors indicate that standard 70% were used for training, while other 30% of the data was used for validation and testing. I wonder how this data was temporally divided onto the training / testing data sets? In my opinion it is important to indicate what is the typical duration of continuous time interval used for training.
4) According to the table 2 it seems that the accuracy decrease nonlinearly increase with increase of prediction time (15, 30, 60 minutes). I think the manuscript would benefit if the authors plot some of the error metrics as a function of prediction time. Further discussion of possible prediction time limits would be also fruitful.
5) According to the figure 3 (by the way, please indicate the X-axis label) certain periodicity of the traffic flow can be observed. Moreover, I think it is quite straightforward that at some temporal scales noticeable periodical repeatable patterns of traffic flow can be found (at least 24-hours pattern obviously should exist in such human-based parameter as traffic flow). Do you think that such a periodicity is actually detected by the proposed traffic flow model? Do traffic flow prediction accuracy in weekdays and holidays are the same?
Author Response
请参阅附件
Author Response File: Author Response.docx
Reviewer 3 Report
The paper deals with the problem of predicting traffic flows in cities. This is a challenging issue and Authors proposed a new way to do it using modeling framework STN-GCN for spatial-temporal normalized graphical convolutional neural networks. The normalization allowed to divide data into high-frequency and low-frequency ones.
The paper consists of six parts. The first one is traditional introduction showing how problematic can be efficient predicting of the discussed topic. It also showed the ways of modeling which were used for forecasting purpose. Therefore, through state-of-the-art analysis concerning methods this chapter introduces the problem in a smooth way explaining also way the proposed approach can be efficient.
The further review of related works was done in second part. Here, graph neural networks, temporal dependence and spatial correlation were discussed with additional citing of several references. Together with Introduction it is a nice background for the next chapters.
Third part presents the concept and structure of STN-GCN model which was a combination of spatial-temporal normalized data types and certain mechanism of transformation. Here, definition of the traffic road network graph was provided as well feature matrix. From the editorial point of view – Figure 1 was pasted here twice – but it can be easily removed. In my opinion, also the explanation of the applied features should be more precise as this subchapter presents only the general approach without information really used in the modeling process.
Next chapter is dedicated to methodology. First, general model framework Is shown consisting of a temporal convolution module STT-block and a spatial-transformer module for the extraction of spatial-temporal features. Again, some editorial mistakes can be observed here, including repeats, bad phrase constructions etc. It must be read once again carefully and some phrases must be changed so the text will be smooth and easily understandable. Next, temporal extraction module is described including normalization procedure and spatial extraction module. The latter was divided into subparts showing spatial location embedding layer, static graph convolution layer, dynamic graph convolution layer and gating mechanism for feature fusion.
All of the stages described in chapter four was the basis for the experiments which were presented in fifth part. First, the setup for it was described, then data was described. Next, evaluation indicators were found including several error measures and all of it lead to the experimental results. The comparison of the predicted and actual values in 60 minutes occurred to be well fitted. Therefore, the proposed methodology can be used efficiently for the forecasting purpose.
The paper ends with conclusions.
In general, the paper is good, but it requires some adjustments. Especially, the editorial aspect needs to be checked and corrected. Some phrases seem to be odd and there are several places where there are repeats of content.
After a minor revision the paper will be good to be published.
In my opinion some phrases are not properly constructed. I think that additional verification is required. Also, there are some phrases with repeated parts. This must be rewritten.
Author Response
Please see the attachment
Author Response File: Author Response.docx
Reviewer 4 Report
The paper reviews aspects of traffic flow forecasting and suggests solutions to enhance accuracy and efficiency by a spatial-temporal normalized graph convolutional neural network model, which can eliminate noises and extract spatial features from the input data.
The paper is nice and I enjoyed reading it; however, I have several concerns:
1. In section 3.1, what is "R^NXN"?
2. Figure 1 appears twice in page 4 and page 5.
3. Figure 1 contains only nodes and edges (V and E). Where is the example for adaptive adjacency matrix A?
4. In Figure 1, G is not defined.
5. The authors write that Delta represents the period of time during which the low-frequency elements remain approximately constant; however, "approximately constant" is not so distinct. It would be better to be more specific. What does the computer do?
6. Sometimes the authors use subscript for the pair (local, global) and the pair (high, low) and sometimes they use superscript. An explanation why using subscript or superscript will be useful.
7. In equation 13, what does the -1/2 mean and how is it combined within the equation? The equation is unclear and there is almost no explanation for the equation and specifically for the constant -1/2.
8. The equations and the results seem to be detached. Please explain how you have designed the experiments based on the theoretical background.
9. The authors write that they optimize and filter the high-frequency elements (both global and local) from the mixed signal. In Li, W.; Li, Z.; Jiang, W.; Chen, Q.; Zhu, G. and Wang, J., "A New Spatial Filtering Algorithm for Noisy and Missing GNSS Position Time Series Using Weighted Expectation Maximization Principal Component Analysis: A Case Study for Regional GNSS Network in Xinjiang Province", Remote Sensors, 2022, Vol. 14, paper no. 1295. Available online at: https://www.mdpi.com/2072-4292/14/5/1295 and also in Rakhmanov, A.and Wiseman, Y., "Compression of GNSS Data with the Aim of Speeding up Communication to Autonomous Vehicles", Remote Sensors, 2023, vol.15, paper no. 2165. Available online at: https://www.mdpi.com/2072-4292/15/8/2165 , the authors suggest filtering out some of the positioning information. I assume that not everything can be filtered out; however, I would encourage the author to cite these papers and explain what other filtering can be done in a possible future work.
Some sentences should be fixed like "In addition, spatial-temporal normalization in introduced to process the stationarity of data" should be "In addition, spatial-temporal normalization is introduced to process the stationarity of data".
Author Response
Please see the attachment
Author Response File: Author Response.docx
Round 2
Reviewer 4 Report
The paper is very similar to the previous version and there are very few changes and the authors have hardly improved the paper. For example in note 9 instead of giving at least an indication of how the filtering out proposed by the papers I proposed would affect, they preferred to simply ignore the note.
Author Response
Please see thPlease see the attachmente attachment.
Author Response File: 
Author Response.pdf
Round 3
Reviewer 4 Report
The authors suggest adding the references as a future work, so that will give an indication how the filtering out proposed by the papers can be applied. This indeed can be a good idea, so I reread the new version of the paper but did not see that the authors added a future work to their paper. They have to take the answer they wrote in the cover letter and put it in the paper as a future work.
