Hierarchical Spatial-Temporal Neural Network with Attention Mechanism for Traffic Flow Forecasting

Round 1

Reviewer 1 Report

I propose to move the descriptions of the y-line in the Figure 3. Extend the conclusions.

Author Response

Thanks to the reviewer's suggestion, on the Figure4(Figure 3 for previous version) , adjusted the distance between the Y-axis title and the scale. And expanded the conclusions, marked in red.

Reviewer 2 Report

This manuscript introduces a novel Hierarchical Spatial-Temporal Neural Networks with Attention Mechanism model (HSTAN). This model concurrently captures temporal correlations and spatial dependencies using a multi-headed self-attention mechanism in both temporal and spatial terms. The paper is well-structured, containing some interesting and useful information. The authors provide a complete framework on the subject, including a detailed description of the methodology. The results are also well-founded and validated, and the conclusions are well-presented.

- The quality of the study would be further improved if a paragraph were added in the concluding section to discuss the added value to the research community and more in-depth discussions on future implications for policy and practice. Besides, future avenues for further studies are encouraged.

Author Response

Thanks to the reviewer's suggestion, on the conclusions, a paragraph were added in the concluding section, marked in red. We plan to investigate the impact of attention on the spatial-temporal graph with a dynamic topology.

Reviewer 3 Report

The presented article is really valued and well prepared. It contains a well-presented mathematical approach and the goals of the algorithm. I only have hesitations about the presentation of the results (Chapters 4.4. and 4.5).

Ad 4.4 The authors present a large amount of information comparing the results of the analysis for various computational models. Presentation in this form (text) does not allow for efficient analysis of the results and drawing conclusions. I would suggest an attempt to organize and visualize these results.

Ad 4.5 The authors decided to visualize the results for the HSTAN and STGCN models. There is no justification why the STGCN model was selected for comparison. It seems that the A3TGCN model obtained results closer to the HSTAN model. Two sentences of commentary on the choice and a reference to Ground Truth would be greatly appreciated. By the way, in chapter 4.1, the scope of data analysis is mentioned for 1-31.01.2015 in the case of SZ-taxi data, and 1.02.20215 appears in the analysis.

Author Response

Thanks to the reviewer's suggestion, 1.02.2015 appears in the Figure 5(Figure 4 for previous version) is a mistake, and we have modified it. In Ad4.4, we added Figure 3 to enhance the visualization of the experimental results in Table 1. The comparison of the experimental results of the HSTAN model with other baseline models at different steps can be better observed through Figure 3. In Ad4.5, The reason for visualizing the actual predictions of the HSTAN and STGCN models is that both models use stacked hierarchies and spatial-temporal blocks. Whereas the A3TGCN model uses an attention mechanism, the overall framework of the model is different from that of HSTAN. We have added relevant statements, marked in red.