Multi-Scale Residual Depthwise Separable Convolution for Metro Passenger Flow Prediction
Round 1
Reviewer 1 Report
The study proposes a neural network approach to the problem of passenger flow determination in the metro based on historical information and graph representation of stations. Despite the detailed description of the methods used in Section 2, this work has a number of disadvantages:
In Section 2 of the work, a formal definition of passenger flow is not presented.
The figures do not show the dimensions of the neural network layers.
There is no the "T" designation on Page 3 lines 112-122 . The definition of the notations ?+1, ?+2, etc. is not given.
Page 7, lines 266-267 refer to Equation 6. However, this equation is not a mathematical explanation of the functionality of Graph Convolutional Network (GCN).
Figure 6 requires further explanation.
Section 3.1. There are no quantitative characteristics of the dataset. It is not justified why data for January 2019 was taken. There is no analysis of metros of other cities. It is not explained what the «status» field in the dataset means.
Page 11, lines 360-361 describe characteristics of the dataset, although table 2 provides a description of hyperparameters.
Table 2 does not reflect all hyperparameters of the model. For example, there are no parameters for the FC layers that are located after RDSC modules, and GCN parameters (MDSC module) are also not presented, see Figure 1. It is not described how these hyperparameters were selected.
The results of the work provide a comparison with baseline models. However, there is no comparison with similar studies on the topic of the problem, which were mentioned in Section 1.
Taking into account the abovementioned disadvantages, the study should be reconsidered after major revision.
Author Response
We have provided a point-by-point response to the reviewer’s comments in the attachement.
Author Response File: 
Author Response.docx
Reviewer 2 Report
This paper addresses an interesting and challenging problem in metro systems management and control. In particular, the authors aim at define a novel model to predict passengers' flow not only at a single station but considering also the correlation with other stations along the line. To this aim the authors present a multi-scale residual depthwise separable convolution network to predict passengers' flows.
The state of the art is sufficiently detailed, however it lacks a discussion on how the passengers' flow prediction can be suitably used in metro traffic control. Works in this regards to be considered are for example: Demand-Oriented Rescheduling of Railway Traffic in Case of Delays, A Service-Oriented Metro Traffic Regulation Method for Improving Operation Performance, Railway disruption: A bi-level rescheduling algorithm, Energy-saving metro train timetable rescheduling model considering ATO profiles and dynamic passenger flow, Distributed approximate dynamic control for traffic management of busy railway networks, Algorithm and peer‐to‐peer negotiation strategies for train dispatching problems in railway bottleneck sections. Additionally, the inclusion of a table summarizing the various contributions of the reviewed papers and the ones of the presented work would be helpful in properly highlighting the contribution of this work with respect to the state of the art and make a clear comparison.
In section 2.1, the intuitive meaning of the correlation graph should be provided. In other words, what does it mean that two stations are correlated? Evidently it does not only represents the connections between adjacent stations. from the Moreover, given that the values of the adjacency matrix elements are either 1 or 0, why do the authors state that the definition set is R? The same comments hold for the similarity graph.
In section 2.2 it is not clear if the correlation of the stations is meant in terms of track connections or also in terms of platforms connections as it happens in complex networks where multiple lines intersect and passengers move from one line to another. Although the authors refer to inter-stations connections the absence of information about the intra-station (i.e., among platforms) passengers flow might provoke an inappropriate prediction.
In section 2.3 the units of figure 2 must be provided.
Section 3 is clearly presented. Since it is necessary to evaluate the effectiveness of the performance of the proposed method it should be clearly specified over which data the method is trained and over which it is tested and compared with the other methodologies.
Typos: The correct name is Hadamard, not Hadamar.
The English language is fine.
Author Response
We have provided a point-by-point response to the reviewer’s comments in the attachement.
Author Response File: Author Response.docx
Round 2
Reviewer 1 Report
All specified disadvantages have been corrected. The work should be accepted in present form.
Reviewer 2 Report
The paper has been sufficiently revised.
