Multi-Task Learning-Based Traffic Flow Prediction Through Highway Toll Stations During Holidays

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

This study introduces a multitask learning framework that combines a Bi-ConvLSTM with a Cross-Attention module to jointly predict the toll station inflow and outflow aiming at the holiday traffic flow. Three year data at Shandong province was applied to test the performance of the proposed strategy. The following problems should be considered.

（1）The innovation point of this study should be further highlighted. The test data, i.e., the holiday traffic flow, is not innovative points, while detailed difference between the proposed algorithm with existing methods should be more clear.

（2）In the simulation study, the data of previous 5 days of May 1^stis applied. Please state the reason. Besides, it would be interesting to study the change of input vector length on the prediction results.

（3）Parameter of the proposed multitask learning framework should be added.

（4）The format of the study should be further improved. For example, In Eq.(1), there is no N and T which is explained in the paragraph below Eq.(1); In Figs. (5) ~ (7), the name of Y-axis is missing.

（5）More graphical results should be added, including the data, the impact of the different parameters of the network on the prediction accuracy, etc.

（6）It is stated that joint prediction of inbound and outbound highway traffic flow are analyzed and predicted, wile in the simulation results, the difference between the inbound and outbound highway traffic flow is not clear. Besides, it will also be interesting to see the impact of the inbound and outbound traffic flow on the prediction results.

（7）There are only several inputs to predict the traffic flow. Is it necessary to apply such complicated conventional network?

Author Response

Thank you very much to Reviewer 1 for carefully reviewing the article. I have attached all the responses from Reviewer 1.

Author Response File: Author Response.docx

Reviewer 2 Report

Comments and Suggestions for Authors

Suggest: 

This paper studies traffic flow prediction at highway toll stations during holidays. The authors propose a multi-task learning-based model for jointly predicting the inflow and outflow at toll stations. The model uses a Spatiotemporal Cross-Attention network to capture the spatial and temporal dependencies of inbound and outbound flow. It further uses a sequence-to-sequence framework to achieve coordinated multi-step prediction. The paper has a clear structure, good logic, and the idea is novel. However, there are still some issues that need to be revised:

1) This paper focuses on long-term traffic flow prediction. Therefore, it would be helpful to briefly introduce some existing long-term traffic flow prediction methods.

2) Since the model uses a multi-task learning framework to jointly predict inbound and outbound flow, the authors are advised to review some related work on multi-task learning. This will help readers better understand the current research in this area.

3) In section Introduction, the paper uses "May Day," while in Section 3.1, it uses "Labor Day." Please use the same term throughout the paper to avoid confusion.

4) The paper uses both “inflow/outflow” and “inbound flow/outbound flow” in different places. Please check the whole text and keep the terminology consistent.

5) There is a problem with the section numbering in section Evaluation. There are two sections labeled as 3.1. Please check and correct the section titles to make sure the numbering is consistent.

6) In section Model settings, please include the experimental configuration, such as computer environment and programming language used.

7) More references are suggested to be incorporated regarding spatiotemporal learning and traffic flow predictions. To name a few,

[1] Yi Z, Zhou Z, Huang Q, et al. Get Rid of Isolation: A Continuous Multi-task Spatio-Temporal Learning Framework[C]//The Thirty-eighth Annual Conference on Neural Information Processing Systems. 2024.

 

Comments on the Quality of English Language

Needs to be some improvement.

Author Response

Thank you very much to Reviewer 2 for his careful review of the article. I have attached all the responses from Reviewer 2.

Author Response File: Author Response.docx

Reviewer 3 Report

Comments and Suggestions for Authors

The paper presents an MTL traffic flow prediction method. It is interesting and may be minor- revised before the final acceptance.

1. The details of the environment and hardware for the algorithms are not given at all. Which version of Python is it？ How about the CPU and GPU？ Please describe all the information. Also, explain how you train the algorithms, do you use Adam? What is the learning rate?
2. How do you divide the data into datasets? What is the ratio of each dataset? What is the parameter of your model? How many layers, how many neurons?
3. The novelty and significance of the present work are not apparent due to the lack of comparison of the results of the references. I highly recommend adding a comparison using data from public datasets used in similar work to clarify the significance.

Author Response

Thank you very much to Reviewer 3 for his careful review of the article. I have attached all the responses from Reviewer 3.

Author Response File: Author Response.docx

Reviewer 4 Report

Comments and Suggestions for Authors

This manuscript proposes a multi-task learning (MTL) approach to predict traffic flow at highway toll stations, particularly under unconventional scenarios, as defined by the authors (e.g., holidays, COVID-19 restrictions). The authors design a novel MTL architecture that decomposes traffic prediction into three tasks: total traffic flow, entrance flow, and exit flow. Using data from 382 toll stations in Jiangsu Province, China, the model is trained and compared against established baselines, including LSTM, GRU, and CNN.

The topic is highly relevant to the intelligent transportation systems (ITS) domain, particularly given the real-world impact of unexpected events on the reliability of traffic predictions. The manuscript is generally well-structured and well-written, with a solid experimental setup and meaningful comparisons. However, there are a few critical issues related to novelty positioning, interpretability, and evaluation robustness that need clarification or improvement. Some major comments are reported below.

1. The authors refer to "unconventional scenarios" as critical cases for prediction, but this concept remains underdefined and insufficiently formalized. It is unclear how these scenarios are identified in practice, or whether they are manually labeled or detected via thresholds. Since the model performance under these scenarios is a central claim, a more rigorous categorization or temporal segmentation of events (e.g., holidays, lockdowns, weather extremes) is necessary.
2. While the multi-task decomposition into entrance/exit/total flow is intuitively sound, the architectural novelty is somewhat limited compared to standard MTL frameworks. The authors should clarify what specific innovations (e.g., shared layer configuration, loss balancing strategy, task dependencies) distinguish this model from prior MTL-based traffic predictors. Also, a qualitative and discursive comparison with other recent deep spatiotemporal or hybrid models (e.g., T-GCN, ST-GAT) would help clarify the incremental contribution.
3. The paper would benefit from a more interpretable breakdown of where and why the MTL model outperforms single-task baselines, particularly in cases involving irregular patterns. For instance, are there toll stations or periods where entrance/exit tasks diverge significantly in difficulty? Including a case analysis or attention map visualization would provide valuable operational insight.
4. The authors use RMSE, MAE, and WMPE  as standard evaluation metrics. However, they should also report standard deviations across multiple runs to reflect model stability. It would be valuable to conduct temporal generalization tests (e.g., training on regular days and testing on holidays) to more accurately assess adaptability to unconventional patterns.
5. More details are needed on data cleaning, handling missing values, and the preprocessing pipeline. For example,  were external variables (e.g., weather, policy restrictions) considered? If not, this should be noted as a limitation, as such features can be crucial in irregular events.

Some minor comments can increase the paper's readability. These are reported below.

• The writing is generally clear but would benefit from light English editing for flow and consistency.
• All acronyms (e.g., MAPE, GRU) should be defined on first use.
• Figures and tables are precise, but Fig. 5, 6, and 7 could be improved with more contrast and labeled axes.

In conclusion, the manuscript presents a valuable and relevant contribution to traffic prediction under non-standard conditions. However, to meet publication standards, the paper requires major revisions.

Author Response

Thank you very much to Reviewer 4 for his careful review of the article. I have attached all the responses from Reviewer 4.

Author Response File: Author Response.docx

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

The comments have been addressed. It is suggested to be accepted and published in the journal.

Reviewer 4 Report

Comments and Suggestions for Authors

Thank you for the extensive revisions made. The revised manuscript presents a more readable and significant contribution. In particular, the redefinition of the scope toward “holiday traffic prediction,” the enhanced explanation of the multi-task learning framework, and the clarifications on data preprocessing and evaluation have significantly improved the manuscript’s clarity, methodological transparency, and practical relevance. The additional experiments and contextual discussion further strengthen the work. In my opinion, the manuscript is now suitable for publication in its current form. Congratulations on your efforts.