Spatio-Temporal Self-Attention Network for Origin–Destination Matrix Prediction in Urban Rail Transit
Round 1
Reviewer 1 Report
Comments and Suggestions for AuthorsThis paper proposes a short-term OD estimation method for urban railways. The suggested method demonstrates improved performance compared to the existing one, making it noteworthy for reporting. I have a few comments for potential improvements.
Major:
1. Lines 20-22 and Lines 66-68: The merit of the proposed method is unclear to me. You mentioned, “…, the results of short-term OD prediction can be used to alleviate congestion during peak hours and improve operational efficiency. …, short-term OD prediction can be used to provide passengers with accurate prediction information.” However, how do you alleviate congestion and operational efficiency using predicted short-term OD? The frequency of railways cannot be easily changed in the short term. Additional explanation is required.
2. Lines 116-117: Is your STSNet a deep learning method? If so, the statement “the proposed STSNet outperforms the advanced deep learning methods (Line 12-13 and other parts)” can be confusing.
3. Lines 134-135: Please clearly specify what is new in your proposed method compared to the existing one. What is your original contribution to the proposed method?
Minor:
1. I suggest removing "termed as" in academic writing (e.g., "OD" and "URT").
2. Line 149: What is “mu_2”?
Author Response
Major:
Comments 1: Lines 20-22 and Lines 66-68: The merit of the proposed method is unclear to me. You mentioned, “…, the results of short-term OD prediction can be used to alleviate congestion during peak hours and improve operational efficiency. …, short-term OD prediction can be used to provide passengers with accurate prediction information.” However, how do you alleviate congestion and operational efficiency using predicted short-term OD? The frequency of railways cannot be easily changed in the short term. Additional explanation is required.
Response 1: Thank you for your valuable comments. In order to make a more accurate and clear description for the advantages of short-term OD prediction, we have modified the description of lines 20-22 in the revised manuscript. For the specific modification, please see the yellow highlighted parts of lines 18-27 in the revised manuscript. In addition, considering that the description of lines 66-68 is not specific and clear, we have already removed the description of lines 66-68 in the revised manuscript.
Comments 2: Lines 116-117: Is your STSNet a deep learning method? If so, the statement “the proposed STSNet outperforms the advanced deep learning methods (Line 12-13 and other parts)” can be confusing.
Response 2: Thank you for making this valuable comment. The proposed SSNet in the revised manuscript is a deep learning method, where we change the name of the proposed method from STSNet to SSNet to make the expression more concise. The meaning of " the proposed STSNet outperforms the advanced deep learning methods" is as follows: At present, for OD prediction in urban rail transit, compared with conventional methods and machine learning methods, deep learning methods have further improved the prediction accuracy and prediction speed. This manuscript uses several existing advanced deep learning methods as the comparison methods to show the superiority of our method. In other words, the three comparison methods (i.e., ConvLSTM, STResNet and CASCNN) are existing advanced deep learning methods, and the proposed SSNet is the latest deep learning method proposed in this manuscript. This manuscript demonstrates the superiority of the proposed SSNet in prediction performance by comparing the proposed SSNet with these three existing advanced deep learning methods. In the revised manuscript, " the proposed STSNet outperforms the advanced deep learning methods" has been modified to “the proposed SSNet outperforms three advanced deep learning methods”. For the specific modification, please see the yellow highlighted parts of lines 11-12 in the revised manuscript.
Comments 3: Lines 134-135: Please clearly specify what is new in your proposed method compared to the existing one. What is your original contribution to the proposed method?
Response 3: Thanks for your constructive comment. We have described the original contribution of the proposed method more clearly by adding the section 2.3. According to section 2.3, for the proposed SSNet, our original contribution is to propose a lightweight yet effective spatio-temporal self-attention module to capture complex long-range spatio-temporal dependencies, thus helping improve the prediction accuracy of the proposed SSNet.
Minor:
Comments 1: I suggest removing "termed as" in academic writing (e.g., "OD" and "URT").
Response 1: Thank you for your valuable comment. We have removed all the "termed as" in the revised manuscript.
Comments 2: Line 149: What is “mu_2”?
Response 2: Thank you for making this valuable comment. The meaning of “mu_2” is as follows: in the revised manuscript, Figure 1 clearly shows the temporal relationship among the historical OD matrices on previous days and the predicted OD matrix. As can be seen from Figure 1, assuming that the time interval of the predicted OD matrix is at the time interval t on day d. Thus, the time interval of the historical OD matrices on previous days is at the time intervals t, t-1, t-2, ... , t-[(mu_2)-1] on each day of the previous mu_1 days. In other words, each day of the previous mu_1 days uses several adjacent OD matrices, specifically, it not only uses the OD matrix at time interval t, but also uses the OD matrices at time intervals t-1, t-2, ... , t-[(mu_2)-1]. Therefore, mu_2 refers to the number of the historical OD matrices used on each day of the previous mu_1 days.
Reviewer 2 Report
Comments and Suggestions for AuthorsThe authors present a very interesting topic through their work with the title: “Spatio-temporal Self-attention Network for Origin-Destination Matrix Prediction in Urban Rail Transit”. My comments for this manuscript are:
1.The authors should include a new sub-section in section 2 and denote their contribution to the existing literature.
2.Based on the literature what other methods have been developed so far for the OD matrix prediction? What are the differences with proposed method and the advantages and disadvantages?
3.The authors should present the historical OD matrices. The historical OD matrices are referring to hours or days? What is the deviation between the historic OD matrices?
4.In the manuscript it is mentioned that the OD dataset contains 50 days. Does it contain weekends? If, Yes then there must be a declination in the entire collected sample.
5.The authors should mention and discuss the limitations of their method.
6.The validation (calibration) of the proposed method is not adequately described.
Author Response
Comments 1: The authors should include a new sub-section in section 2 and denote their contribution to the existing literature.
Response 1: Thank you for your valuable comments. We have added a new subsection (i.e., Section 2.3), and added our contribution to the existing literature in Section 2.3. For the specific additions, please see the Section 2.3.
Comments 2: Based on the literature what other methods have been developed so far for the OD matrix prediction? What are the differences with proposed method and the advantages and disadvantages?
Response 2: Thank you for making this valuable comment. Due to the difficulty of obtaining a large number of passenger flow data of urban rail transit, there are few short-term OD matrix prediction methods based on deep learning in urban rail transit. To the best of our knowledge, the proposed method in this manuscript may be the first method to introduce a lightweight self-attention module into short-time OD matrix prediction of urban rail transit. At present, we have not found the latest research on the short-term OD matrix prediction method of urban rail transit based on deep learning that is related to the literature in Section 2.1 and Section 2.2.
Comments 3: The authors should present the historical OD matrices. The historical OD matrices are referring to hours or days? What is the deviation between the historic OD matrices?
Response 3: Thanks for your constructive comments. We have presented the historical OD matrices in Figure 1 of the revised manuscript. According to Figure 1, the historical OD matrices on previous mu_1 days are used as the inputs. Assuming that the time interval of the predicted OD matrix is at the time interval t on day d. Thus, each day of the previous mu_1 days uses several adjacent OD matrices, specifically, each day of the previous mu_1 days not only uses the OD matrix at time interval t, but also uses the OD matrices at time intervals t-1, t-2, ... , t-[(mu_2)-1]. Besides, the time granularity of the data we have obtained in this manuscript is 5 minutes. Therefore, in this manuscript, the time deviation between adjacent historical OD matrices in the same day is 5 minutes, and the time deviation between historical OD matrices on the same time interval of adjacent dates is 1 day.
Comments 4: In the manuscript it is mentioned that the OD dataset contains 50 days. Does it contain weekends? If, Yes then there must be a declination in the entire collected sample.
Response 4: Thank you for your valuable comment. The OD dataset used in the manuscript contains weekends. We use the OD dataset from weekends for the following reasons: firstly, using the OD dataset from weekends can make full use of the obtained weekend data and thus improve the utilization of obtained OD dataset. Secondly, if only weekday data are used, overfitting is easy to occur in the training process because the data of each day on weekdays usually follow the same distribution. If a few weekend data are added, since the data distribution on weekends is usually different from that on weekdays, the diversity of data can be improved, and thus the occurrence of overfitting can be better prevented.
Comments 5: The authors should mention and discuss the limitations of their method.
Response 5: Thank you for making this valuable comment. We have added the limitations of our method in the conclusion section of the revised manuscript. For the specific additions, please see the yellow highlighted parts from line 530 to line 532 in the revised manuscript. The main limitations of the proposed method can be summarized as follows: the proposed SSNet has only one input, so the proposed SSNet is a deep learning method based on single-source data, which can only use OD data as input, and cannot use other multi-source data (such as weather data and mobile phone data) as input.
Comments 6: The validation (calibration) of the proposed method is not adequately described.
Response 6: Thanks for your constructive comment. Our understanding for this comment is that we did not adequately describe the convergence of the validation results of the proposed method in the training process in the original manuscript. Based on such understanding, we add the comparisons of convergence curves of best validation RMSE among the proposed SSNet and the three comparison methods in Figure 9, and the comparison results among the proposed SSNet and the comparison methods are further analyzed in the yellow highlighted parts from line 440 to line 443 in Section 4.5.1 of the revised manuscript.
Reviewer 3 Report
Comments and Suggestions for AuthorsThe objective of this work is the proposal of a new methodology for short-term origin-destination matrix for improving the prediction accuracy in urban rail transit operation. The proposed model was compared to previous studies on this regard in a case study. There are mistakes in the Figure 1 that must be corrected. On the other hand, RMSE abbreviation was used (lines 344 and 345) before the complete explanation of the parameter (line 349). In addition to all that, the main aspect that needs to be reviewed is the lack of detailed practical applications to improve the operation of urban rail transit from the point of view of sustainability, which is the scope of the journal, with the inclusion of adequate references related to this specific issue.
Comments on the Quality of English LanguageThe text is well written and the English is reasonably good, but some typing errors were found, especially regarding the use of capital letters, which need to be corrected.
Author Response
Comments 1: The objective of this work is the proposal of a new methodology for short-term origin-destination matrix for improving the prediction accuracy in urban rail transit operation. The proposed model was compared to previous studies on this regard in a case study. There are mistakes in the Figure 1 that must be corrected. On the other hand, RMSE abbreviation was used (lines 344 and 345) before the complete explanation of the parameter (line 349). In addition to all that, the main aspect that needs to be reviewed is the lack of detailed practical applications to improve the operation of urban rail transit from the point of view of sustainability, which is the scope of the journal, with the inclusion of adequate references related to this specific issue.
Response 1: Thank you for your valuable comments.
For your first question, in the revised manuscript, we have further improved Figure 1 to make the relationship among the historical OD matrices on previous days and the predicted OD matrix more intuitive. According to the improved Figure 1 in the revised manuscript, assuming that the time interval of the predicted OD matrix is at the time interval t on day d. Thus, the time interval of the historical OD matrices on previous days is at the time intervals t, t-1, t-2, ... , t-[(mu_2)-1] on each day of the previous mu_1 days. We currently find no errors that need to be corrected.
For your second question, we have added the complete explanation of RMSE where it first appears in the revised manuscript. For the specific additions, please see the yellow highlighted parts from line 358 to line 359 in the revised manuscript.
For your third question, in the revised manuscript, we have analyzed the significance of short-term OD prediction for improving the operation of urban rail transit from the point of view of sustainability, please see the yellow highlighted parts from line 18 to line 27 in the revised manuscript. Then, we have added relevant literature and a description of the literature from the point of view of sustainability. For the specific additions, please see Ref. [19] and the yellow highlighted parts from line 112 to line 118 in the revised manuscript.
Comments 2: The text is well written and the English is reasonably good, but some typing errors were found, especially regarding the use of capital letters, which need to be corrected.
Response 2: Thank you for making this valuable comment. Some typing errors and other language issues have been checked and corrected in the revised manuscript. For specific modifications, please see the yellow highlighted parts in lines 28, 70, 429 of the revised manuscript.
Round 2
Reviewer 1 Report
Comments and Suggestions for AuthorsThank you for your revision. I'm satisfied.
Reviewer 2 Report
Comments and Suggestions for AuthorsThe authors have addressed all my comments.
Reviewer 3 Report
Comments and Suggestions for AuthorsThe suggestions presented for the review of the article were mostly appreciated, so that the article may be accepted for publication in the current version.
