Parallel Interactive Attention Network for Short-Term Origin–Destination Prediction in Urban Rail Transit

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

M/s: Parallel Interactive Attention Network for Short-term Origin-Destination Prediction in Urban Rail Transit

In this paper, a novel parallel interactive attention network (termed PIANet) for short-term origin-destination prediction in urban rail transit is proposed to further improve the short-term OD prediction accuracy. In the proposed PIANet, a novel omnidirectional attention module (termed OAM) is proposed to improve the representational power of the network by calculating the feature weights in both channel axes and spatial dimensions. Besides, a simple yet effective feature interaction is proposed to improve the feature utilization. Based on the two real-world datasets from the Beijing subway, the comparative experiments demonstrate that the proposed PIANet outperforms the state-of-the-art deep learning methods for short-term OD prediction in URT, and the ablation studies show that the proposed OAMs and feature interaction play an important role in improving the short-term OD prediction accuracy.

My comments are as follows:

1. Please list the contributions of the paper after the related work subsection. Further, first, provide the complete detail of the acronym, for example, OD in the abstract.

2. Include a pseudo code that describes the sequence of the steps of the proposed method.

3. What is the significance of error-based indicators such as RMSE, MAE, and SMAPE? Why not use any one of them to test the accuracy of the model? Discuss.

4. The description of the data set used in terms of the range of individual variables must be described and highlighted properly. Is any pre-processing done? What is the difference between validation and test data? Discuss in detail.

5. Is the proposed model which is quite complex required for the used data (since the amount of data appears to be quite small). Discuss.

6. The learning algorithm novelty needs to be properly discussed as what the contributions of the authors in its development. Discuss.

7. Various recent ANN prediction models such as Memory Recurrent Elman Neural Network-Based Identification of Time-Delayed Nonlinear Dynamical System etc. & other related papers are available in the literature whose discussion is missing from the paper.

8. The paper also has some minor typos and language issues, which must be checked and corrected in the revision.

I suggest a revision.

Comments on the Quality of English Language

The paper also has some minor typos and language issues, which must be checked and corrected in the revision

Author Response

Comments 1: Please list the contributions of the paper after the related work subsection. Further, first, provide the complete detail of the acronym, for example, OD in the abstract.

Response 1: Thank you for your valuable comments. In the related work section of the revised manuscript, we add the contributions of this paper at the end of Section 2.1. For the specific additions, please see the yellow highlighted parts from line 111 to line 118 in the revised manuscript.

Besides, in the revised manuscript, we have added the complete details of the acronyms where the acronym first appears. For example, in line 1 of the abstract, “OD” is changed to “origin-destination (termed as OD)” where OD first appears.

Comments 2: Include a pseudo code that describes the sequence of the steps of the proposed method.

Response 2: Thank you for making this valuable comment. In our revised manuscript, we have added the pseudo code that describes the sequence of the steps of the proposed PIANet. For the specific additions, please see Table 1 in the revised manuscript.

Comments 3: What is the significance of error-based indicators such as RMSE, MAE, and SMAPE? Why not use any one of them to test the accuracy of the model? Discuss.

Response 3: Thanks for your constructive comments. The answer to your first question is: According to Eq. (20) and Eq. (21) in the revised manuscript, RMSE and MAE are two different evaluation metrics that can measure the absolute error between the predicted OD matrix and the real OD matrix, but cannot reflect whether the predicted OD matrix is underestimated or overestimated relative to the real OD matrix. According to Eq. (22) in the revised manuscript, SMAPE as a relative error metric can reflect whether the predicted OD matrix is underestimated or overestimated relative to the real OD matrix. Therefore, the prediction performance of the network can be comprehensively evaluated by using RMSE, MAE, and SMAPE. The significance of RMSE, MAE and SMAPE has been added in the revised manuscript. For the specific additions, please see the yellow highlighted parts from line 316 to line 322 in the revised manuscript.

The answer to your second question is that we have used RMSE, MAE, and SMAPE to test the accuracy of the model in section 4.2. For example, RMSE, MAE, and SMAPE in Table 3 and Table 4 in the revised manuscript are all used as the accuracy of the model.

Comments 4: The description of the data set used in terms of the range of individual variables must be described and highlighted properly. Is any pre-processing done? What is the difference between validation and test data? Discuss in detail.

Response 4: Thank you for your valuable comments. I'm sorry to tell you that we don't understand what " the range of individual variables " means. However, we have modified the description of the data set to make the description of the data set clearer. For specific modifications, please see the yellow highlighted parts in section 4.1.1 of the revised manuscript.

We used the Min-Max normalization for preprocessing of the input data, and we have added a relevant description of the preprocessing in the revised manuscript. For the specific additions, please see the yellow highlighted parts from line 301 to line 303 in the revised manuscript.

The difference between the validation data and the test data is as follows: The validation data is used in the training process to evaluate how well the neural network model is trained. Specifically, it is common to train a neural network model for multiple epochs, but after each epoch, it is difficult to know how well the training has been done, so the model performance can be evaluated by using the validation data after each epoch to see how well the neural network model was trained. If it is found that the convergence of the neural network model in the training process is not good, the training can be stopped in time to seek the reason for the poor convergence in time, which can greatly save computational cost and time cost. After the neural network model is trained, in the testing process, the test data is used to evaluate the final performance of the trained neural network model.

Comments 5: Is the proposed model which is quite complex required for the used data (since the amount of data appears to be quite small). Discuss.

Response 5: Thank you for making this valuable comment. We think that the complexity of the proposed model is matched with the amount of data, and the proposed model can achieve high-precision short-time OD prediction with the used data. The main reasons are as follows:

For BJSubway2021 and BJSubway2022, the number of data records reaches 137 million and 98 million respectively, and the data records in BJSubway2021 and BJSubway2022 can be converted into 10200 and 8976 OD matrices respectively. It can be seen that the amount of data in BJSubway2021 and BJSubway2022 is not quite small. Besides, many OD matrices are sparse, and there are complex spatio-temporal correlations among different OD matrices, therefore, it is difficult to obtain high prediction accuracy if a simple model is used. For example, PIANet-D removes the feature interaction in the PIANet architecture and thus can be viewed as a simplified variant of PIANet. According to Table 4 in the revised manuscript, the prediction performance of PIANet-D decreases compared to PIANet.

Comments 6: The learning algorithm novelty needs to be properly discussed as what the contributions of the authors in its development. Discuss.

Response 6: Thanks for your constructive comment. The learning algorithm novelty mentioned in the manuscript contains two aspects: (1) a novel omnidirectional attention module (termed as OAM) is proposed to improve the representational power of the network; (2) a simple yet effective feature interaction is proposed to improve the feature utilization. We have added more discussion about OAM and feature interaction in the introduction section of the revised manuscript. For the specific additions, please see the yellow highlighted parts from line 54 to line 63 in the revised manuscript.

Comments 7: Various recent ANN prediction models such as Memory Recurrent Elman Neural Network-Based Identification of Time-Delayed Nonlinear Dynamical System etc. & other related papers are available in the literature whose discussion is missing from the paper.

Response 7: Thank you for your valuable comment. In the revised manuscript, We have added three related papers in the references section, and appropriate discussion in the related work section. For the specific additions, please see the yellow highlighted parts from line 504 to line 509 and from line 99 to line 110 in the revised manuscript.

Comments 8: The paper also has some minor typos and language issues, which must be checked and corrected in the revision.

Response 8: Thank you for pointing this out. Some minor typos and language issues have been checked and corrected in the revised manuscript. For specific modifications, please see the yellow highlighted parts in lines 24, 43, 129, 244, 245, and 327 of the revised manuscript.

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

Deep learning methods have gained significant attention in recent years. Therefore, it is worth to investigate the application methods of deep learning for URL OD prediction. the paper is well organised. However there are some ponints that need tobe corrected. Please find the details as given follows:

- In section 2.1, there are some mentioned former studies. Please include their outcomes, weaknesses of the algorithms, and recommendations. This will improve the novelty of your research.

- In the discussion section, the results should be discussed in detail including the knowledge in the literature.

Comments on the Quality of English Language

Please do a proof reading.

Author Response

Comments 1: In section 2.1, there are some mentioned former studies. Please include their outcomes, weaknesses of the algorithms, and recommendations. This will improve the novelty of your research.

Response 1: Thank you for your valuable comments. In section 2.1 of the revised manuscript, we have added the outcomes of some former studies, and have also added the weaknesses of some former studies. For the specific additions, please see section 2.1 in the revised manuscript.

Comments 2: In the discussion section, the results should be discussed in detail including the knowledge in the literature.

Response 2: Thank you for making this valuable comment. In the discussion section of the revised manuscript, the results have been discussed including the knowledge in the literature in lines 386 to 399 of the revised manuscript.

Comments 3: Please do a proof reading.

Response 3: Thank you for pointing this out. Some minor typos have been checked and corrected in the revised manuscript. For specific modifications, see the yellow highlighted parts in lines 24, 43, 129, 244, 245, and 327 of the revised manuscript.

Author Response File: Author Response.pdf

Reviewer 3 Report

Comments and Suggestions for Authors

The article addresses a critical aspect of urban rail transit operations by proposing a novel approach for short-term Origin-Destination (OD) prediction. Research provides a well-motivated, well-structured, and thoroughly researched exploration of short-term OD prediction in urban rail transit. The proposed PIANet, with its omnidirectional attention module and effective feature interaction, demonstrates promising results.

 

1. Considering the diverse nature of urban rail transit systems, it would be insightful to explore how the PIANet framework could be adapted or extended to different contexts. It would be beneficial to discuss the potential applicability of PIANet beyond the specific context of the Beijing subway. Consider addressing whether the proposed approach could be adapted or extended to other urban rail transit systems with different characteristics or scales.

2. The conclusion could be strengthened by summarizing the key findings and reiterating the practical implications of PIANet on short-term OD prediction in urban rail transit. Emphasizing the broader significance of the proposed solution in addressing real-world challenges would enhance the paper's conclusion.

3. The theme of short-term OD prediction with neural networks is experiencing a boom at the moment. It seems that there are new publications coming every month.

Although the research is absolutely correct and results show high accuracy, the research can be improved if the review section is expanded with new articles on this theme, for example:

 

Yang, Fang, et al. "Predictability of short-term passengers’ origin and destination demands in urban rail transit." Transportation 50.6 (2023): 2375-2401.

He, Yuxin, Yang Zhao, and Kwok-Leung Tsui. "Short-term forecasting of origin-destination matrix in transit system via a deep learning approach." Transportmetrica A: Transport Science 19.2 (2023): 2033348.

Zhang, Yan, et al. "Deep Learning for Metro Short-Term Origin-Destination Passenger Flow Forecasting Considering Section Capacity Utilization Ratio." IEEE Transactions on Intelligent Transportation Systems (2023).

Dong, Ningning, et al. "A method for short-term passenger flow prediction in urban rail transit based on deep learning." Multimedia Tools and Applications (2023): 1-23.

4. The "three-dimensional space" term is somewhat confusing because it implies the spatial metrics, width, height and length or x,y,z dimensions. Maybe a proper substitute for this term can be found in your research?

Author Response

Comments 1: Considering the diverse nature of urban rail transit systems, it would be insightful to explore how the PIANet framework could be adapted or extended to different contexts. It would be beneficial to discuss the potential applicability of PIANet beyond the specific context of the Beijing subway. Consider addressing whether the proposed approach could be adapted or extended to other urban rail transit systems with different characteristics or scales.

Response 1: Thank you for your valuable comments. The PIANet can be adapted or extended to other urban rail transit systems with different characteristics or scales for the following reasons: In the PIANet, the OAMs can enhance the representational power of the network, and the feature interaction can further improve the feature utilization, which makes the PIANet have excellent feature extraction ability. When the PIANet is extended to another urban rail transit system with different characteristics and scale, only the sizes of input and output need to be modified according to the scale of the new urban rail transit system, and then the corresponding input data and output data are used for training. Due to its excellent feature extraction ability, the PIANet can capture the new passenger flow distribution during training. Therefore, after training, PIANet can adapt to the new urban rail transit system and achieve accurate short-term OD prediction. We have added the extensibility of the PIANet in the second paragraph of the conclusion of the revised manuscript.

Comments 2: The conclusion could be strengthened by summarizing the key findings and reiterating the practical implications of PIANet on short-term OD prediction in urban rail transit. Emphasizing the broader significance of the proposed solution in addressing real-world challenges would enhance the paper's conclusion.

Response 2: Thanks for your constructive comments. In the revised manuscript, the conclusion has been strengthened by summarizing the key findings and reiterating the practical implications of the PIANet on short-term OD prediction in urban rail transit in the first paragraph of the conclusion. Besides, we have added the extensibility of the PIANet as the broader significance of the proposed PIANet in the second paragraph of the conclusion.

Comments 3: The theme of short-term OD prediction with neural networks is experiencing a boom at the moment. It seems that there are new publications coming every month.

Although the research is absolutely correct and results show high accuracy, the research can be improved if the review section is expanded with new articles on this theme, for example:

Yang, Fang, et al. "Predictability of short-term passengers’ origin and destination demands in urban rail transit." Transportation 50.6 (2023): 2375-2401.

He, Yuxin, Yang Zhao, and Kwok-Leung Tsui. "Short-term forecasting of origin-destination matrix in transit system via a deep learning approach." Transportmetrica A: Transport Science 19.2 (2023): 2033348.

Zhang, Yan, et al. "Deep Learning for Metro Short-Term Origin-Destination Passenger Flow Forecasting Considering Section Capacity Utilization Ratio." IEEE Transactions on Intelligent Transportation Systems (2023).

Dong, Ningning, et al. "A method for short-term passenger flow prediction in urban rail transit based on deep learning." Multimedia Tools and Applications (2023): 1-23.

Response 3: Thank you for making this valuable comment. The several new publications you provide are very meaningful for this manuscript, and we have put the related new publications into the references in the revised manuscript, and then cited and discussed them in the related work section. For the specific additions, please see the yellow highlighted parts from line 504 to line 509 and from line 99 to line 110 in the revised manuscript.

Comments 4: The "three-dimensional space" term is somewhat confusing because it implies the spatial metrics, width, height and length or x,y,z dimensions. Maybe a proper substitute for this term can be found in your research?

Response 4: Thank you for pointing this out. In the original manuscript, "three-dimensional space" actually refers to the three-dimensional coordinate space consisting of the channel axis, width axis and height axis. “spatial dimension” refers to the two-dimensional coordinate space consisting of width axis and height axis. To keep the term names consistent, we change "three-dimensional space" to “channel-spatial dimension” in the revised manuscript, besides, the meaning of “channel-spatial dimension” has been clearly described in lines 134 to 135 in the revised manuscript.

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

The authors have satisfactorily responded to my questions except for comment 7 "Various recent ANN prediction models such as Memory Recurrent Elman Neural Network-Based Identification of Time-Delayed Nonlinear Dynamical System etc.", the author overlooked the suggestion and did not comment on how the suggested reference can be applied to the present research?

Author Response

Comment 1: The authors have satisfactorily responded to my questions except for comment 7 "Various recent ANN prediction models such as Memory Recurrent Elman Neural Network-Based Identification of Time-Delayed Nonlinear Dynamical System etc.", the author overlooked the suggestion and did not comment on how the suggested reference can be applied to the present research?

Response 1: Thank you for your valuable comment. In the revised manuscript, We have added the suggested paper named "Memory Recurrent Elman Neural Network-Based Identification of Time-Delayed Nonlinear Dynamical System" in the references section of the revised manuscript. Since the paper named "Memory Recurrent Elman Neural Network-Based Identification of Time-Delayed Nonlinear Dynamical System" is not very relevant to the field of short-time OD prediction in urban rail transit, we put the citation of this paper in the introduction section of the revised manuscript. For the specific additions, please see the yellow highlighted parts in lines 45, 496, and 497 of the revised manuscript.