Large-Scale Station-Level Crowd Flow Forecast with ST-Unet
Round 1
Reviewer 1 Report
The paper proposes a Spatio-Temporal U-shape network model for station-level crowd flow forecast that emphasizes stations’ spatial dependencies by integrating the crowd flow information from neighboring stations.
The paper is difficult to follow. The structure of the paper and the English language can be improved to improve the reader's understanding. There are some clear typos.
Examples after reading the Abstract:
" Forecasting crowd flow for such places, called station-level crowd flow forecast in this paper, would benefit many applications, like traffic management and public safety etc. Concretely, station-level forecast (Comment: is this the same as Station-level crowd flow forecast??) is to forecast (Comment: use predict) the number of people that will arrive at or depart from ‘stations’ in a future period."
"ST-Unet emphasizes stations’ spatial dependencies by integrating the crowd flow information from neighboring stations and belonging clusters clustered by hierarchical clustering (Comment: three clusters)."
Section 2.2
"the receptive field of CNNs of each entry in regular grid data, is its 8 or 24 neighbor grids (3x3 or 5x5 feature maps)." (I can not see the connection between those two concepts)
I would encourage the authors to write the paper in a way that is easy to understand and follow.
Author Response
Dear reviewer,
Thanks for your valuable comments and suggestions, which are very helpful in improving the quality of our work. All the comments have been addressed as follows.
Point 1: The paper is difficult to follow. The structure of the paper and the English language can be improved to improve the reader's understanding. There are some clear typos.
Response 1: The English expressions have been improved by a native English speaker.
Point 2: " Forecasting crowd flow for such places, called station-level crowd flow forecast in this paper, would benefit many applications, like traffic management and public safety etc. Concretely, station-level forecast (Comment: is this the same as Station-level crowd flow forecast??) is to forecast (Comment: use predict) the number of people that will arrive at or depart from ‘stations’ in a future period."
Response 2: The expressions has been revised. We use SLCFF as the abbreviation of “Station-level crowd flow forecast”.
Point3: "ST-Unet emphasizes stations’ spatial dependencies by integrating the crowd flow information from neighboring stations and belonging clusters clustered by hierarchical clustering (Comment: three clusters)."
Response 3: Thank you for your suggestions.
Point 4: "the receptive field of CNNs of each entry in regular grid data, is its 8 or 24 neighbor grids (3x3 or 5x5 feature maps)." (I cannot see the connection between those two concepts)
Response 4: The expression has been revised. For regular grid data, each entry’s receptive field under a 3x3 feature map is its 8 neighbor grids (colored white as follows).
Author Response File: 
Author Response.pdf
Reviewer 2 Report
p.p1 {margin: 0.0px 0.0px 0.0px 0.0px; font: 12.0px Helvetica} p.p2 {margin: 0.0px 0.0px 0.0px 0.0px; font: 12.0px Helvetica; min-height: 14.0px}This paper proposes a deep learning based model, named ST-Unet, to do station-level crowd flow forecast. The combination of geographic information with the design of neural networks is shown. On the whole, three Unet branches to capture the temporal influence and one branch to introduce external influence are integrated into the forecast model. The overall results of ST-Unet outperform baselines on station-level crowd forecast.
The proposed method does not show much superiority in the instance of rain, foggy, and holidays. The author indicated that the potential reasons for that may be the insufficiency of training data. However, we can observe that crowd flows on a rainy day and non-rainy day are obviously different so there could be other factors affecting the performance regarding the instance of rain at least. The author may investigate the algorithm to indicate potential reasons other than the insufficiency of training data.
Author Response
Dear reviewer,
Thanks for your valuable comments and suggestions, which are very helpful in improving the quality of our work. All the comments have been addressed as follows.
Point 1: The proposed method does not show much superiority in the instance of rain, foggy, and holidays. The author indicated that the potential reasons for that may be the insufficiency of training data. However, we can observe that crowd flows on a rainy day and non-rainy day are obviously different so there could be other factors affecting the performance regarding the instance of rain at least. The author may investigate the algorithm to indicate potential reasons other than the insufficiency of training data.
Response 1: Thank you for your comments. We’ve found that most time periods of foggy days in the test datasets are in the evening or midnight, such as 21:00-21:20 on 26/10/2016, 23:00 on 26/10/2016 – 5:30 on 27/10/2016 in Chicago. And the precipitation-accumulation (1 hour) on rainy days are 0.43 millimeter on average and 0.54 millimeter max on 29/10/2017 in New York City. We think the influence on crowd flow is not obvious so that ST-Unet only performs slightly better than other methods.
Author Response File: Author Response.pdf
Reviewer 3 Report
Large Scale Station-Level Crowd Flow Forecast with ST-Unet
------------
Paper STRENGTH: Experiments on four real-world dataset are carried out to verify the proposed method’s performance and the results show that ST-Unet outperforms seven baselines on station-level crowd forecast.
RESEARCH:
Good research ! Congratulations !
Do you have any repository on Git with such research? I will be a great idea to mention it on the paper.
ENGLISH:
Minor revision.
Author Response
Dear reviewer,
Thanks for your valuable comments and suggestions, which are very helpful in improving the quality of our work. All the comments have been addressed as follows.
Point 1: Do you have any repository on Git with such research? I will be a great idea to mention it on the paper.
Response 1: I'd share our code on https://github.com/zhouyirong09/ST-Unet.git this month.
Point 2: ENGLISH: Minor revision.
Response 2: The English expressions have been improved by a native English speaker.
Author Response File: Author Response.pdf
Round 2
Reviewer 1 Report
My comments have been included. The article now is easier to read and understand.
The research itself is very well delivered.
