Attention-Based Multiple Graph Convolutional Recurrent Network for Traffic Forecasting

Round 1

Reviewer 1 Report

The authors presented an interesting paper that introduced an Attention-based Multiple Graph Convolutional Recurrent Network (AMGCRN) for traffic forecasting. The reviewer considers the following points need to be addressed for the acceptance of the paper:  

1)     The reviewer suggests adding numerical results at the end of the abstract.

2)     Some sentences are unclear due to grammatical errors. Please consider proofreading to enhance the readability of the paper.

3)     Figure 1 provides a graphical representation of road nodes visualisation of the PeMSD4 dataset. The legend indicates example of nodes. Please provide further information about these and the nomenclature utilised.

4)     Contributions should address the novelties introduced in the manuscript. Experiments on two real-world traffic datasets is part of the validation process, and thus it should not be included in the contributions section.

5)     The reviewer suggests including the contributions at the end of the related works section, once all the gaps related to traffic forecasting have been identified.

6)     A table summarising the methods introduced in the related works section with the advantages and disadvantages can be helpful to the reader to understand even more the contributions of this manuscript.

7)     Section 2.2. Graph based methods can be expanded including the latest advancements, as the authors only included 7 references for analysis.

8)     The authors only analysed two types of methods in the related works section: traditional methods and graph-based methods. The authors then determined that these methods cannot simultaneously capture global and dynamic local spatio-temporal correlations. As this is the main gap identified, the reviewer considers that the analysis of spatio-temporal networks for traffic forecasting need to be also analysed and discussed. Which is the advantages of the proposed approach if compared with the latest advancements in spatio-temporal networks?

9)     There is no evidence that the generalisation performance is validated. How did you analyse that the generalisation of the proposed approach? Validating the proposed approach through the split of the dataset into distinct sets is not enough.

10)  The selection of the baseline models needs to be justified. Additionally, other state-of-the-art methods need to be included in the analysis if a comparative study is utilised to demonstrate the effectiveness of the proposed method.

11)  Analogously, the metrics utilised are not justified.

12)  As part of the data pre-processing the authors highlighted the need to window and aggregate the data. Did the authors need to deal with challenges such as missing values and outliers? How did the authors address such challenges?

13)  Learning parameters such as optimiser, batch size, and epochs were not justified. How did the authors select their values?

14)  Metrics such as MAE, RMSE, and MAPE were utilise to assess the performance of the models. What about the execution time?

15)  It can be perceived that when considering the PeMSD8 dataset the AGCRN model outperformed AMGCRN. If the PeMSD4 is considered it can be perceived that the difference in performance between AGCRN and AMGCRN is not significant. Why should one consider your approach instead of AGCRN? The results section needs to be expanded to further discuss these results.

16)  Which are the limitations of your approach?

Author Response

Dear Reviewer,

We extend our sincerest thanks for reviewing our paper and providing valuable feedback and suggestions. Your insights have been instrumental in enabling us to enhance the overall quality of our work. Following your suggestions, we have thoroughly revised the paper, incorporating substantial changes to showcase the unique strengths and benefits of our proposed method. We are confident that these revisions will address your concerns and bring our work to a higher standard.Please see the attachment.

Best regards.

Author Response File: Author Response.pdf

Reviewer 2 Report

The authors propose an Attention-based Multiple Graph Convolutional Recurrent Network framework to model complex spatio-temporal correlations in the traffic flow forecasting task. I have some suggestions on this paper.

1- The title is not so good. The main idea of the paper is not reflected in the title.

2- The abstract can be revised to include more details of the method and conclusions.

3- 5.1. Datasets: More datasets are needed to verify the method.

4- Table 2: Compared with Model AGCRN, the advantages of the method proposed by the author are not obvious.

5- The conclusions should grasp the main findings and the results of this study.

Author Response

Dear Reviewer,

We extend our sincerest thanks for reviewing our paper and providing valuable feedback and suggestions. Your insights have been instrumental in enabling us to enhance the overall quality of our work. Following your suggestions, we have thoroughly revised the paper, incorporating substantial changes to showcase the unique strengths and benefits of our proposed method. We are confident that these revisions will address your concerns and bring our work to a higher standard.Please see the attachment.

Best regards.

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

The paper can be accepted.

Reviewer 2 Report

The paper has been carefully revised. My suggestion is to accept it.