RoPT: Route-Planning Model with Transformer

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

The reviewed articel presents an inovative approach based on the fusion of Graph Convolutional Networks (GCN) and Transformer, referred as RoPT. This model was evaluated in relation to real taxi ride routs, and its effectivness was compared with classical methods such as Dijkstra, A*, and others. The authors conducted a reliable comparision of these methods with actual taxi rides, which forms a strong foundation for the analysis. However, I have seems to have an important question: to what extent can real taxi rides be considered as a referential standard, given that they are not necessarily optimal in terms of distance or travel time?

Main consern

The fundamental issue of the articel is the assumption that real taxi routs are a suitable reference point for evaluating the effectiveness of route prediction methods. The authors correctly conducted a comparision of classical methods and the new RoPT approach against real taxi rides, but there was a lack of analysis on how these routes compare to theoretically optimal solutions determined by precisely those classical algorithms.

Consequences of this shortcoming:

• It is unknown whether real taxi routs are optimized in terms of distance and travel time. Drivers may make decisions based on subjective factors, such as knowledge of the city, passenger availability, or even individual preferences, which may distort the model's performance evaluation.
• Lack of reference to classical algorithms as a optimality benchmark. Dijkstra and A* are widely recognized as methods for finding shortest routes, so comparing RoPT’s results also against these optimal routes would provide a more complete picture of the model's effectiveness.
• It is not possible to clearly evaluate the added value of RoPT in the context of rational route planning patterns. The model may better adapt to driver behavior, but does not necessarily provide solutions closer to real optimization

Suggestion for improvement:

1. Extend the analysis to include an evaluation of the consistency of classical methods' results (Dijkstra, A*, etc.) in relation to both taxi routes and theoretically optimal routes.
2. Analyze the differences between real routes and those determined by classical algorithms. Do drivers choose less efficient (what is the objecitve function - optimality?) routes due to dynamic traffic conditions, personal preferences, or infrastructural constraints?
3. Incorporate these findings in the conclusion of the article to provide a fuller context for assessing RoPT's efficiency.

Conclusion

The articel is solidly prepared, provides valuable results, and makes a significant contribution to the field of trajectory prediction, offering a new perspective on combining GCN and Transformer in the problem of route planning. The detailed analysis of real taxi rides is valuable, but to gain a fuller picture of RoPT's effectiveness, it is worth including a comparision with optimal routes determined by classical algorithms. This approach would allow to clearly determine whether RoPT actually finds more optimal routes or merely better adapts to real driver choices.

I recommend including this additional analysis in future research, which will allow for a more comprehensive assessment of RoPT’s potential and its real advantage over classical methods.
It is worth using new comparissons to expand the scarce conclusion section.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

The abstract exceeds 200 words, which does not align with the journal's guidelines. It should be more structured, with clearer distinctions between background, methods, results, and conclusions. The claims should be more balanced—avoiding overstatements about performance improvement while clearly emphasizing contributions.

The introduction provides a solid foundation but lacks a more explicit explanation of the research gap. Additional references to recent studies in route planning using Transformer and GCN would strengthen the contextualization of the study.

While the methodology is well-structured, some implementation details are missing, particularly regarding the hyperparameter selection process and computational efficiency considerations. Including a step-by-step explanation of the RoPT model architecture, particularly the interaction between GCN and Transformer, would improve reproducibility.

The tables and figures are clear, but some captions could be more self-explanatory. In Table 3 and Table 4, bolding key performance improvements would help emphasize RoPT’s advantages over other models.

The results are well presented, but the discussion lacks deeper analysis on limitations. A critical discussion on the potential biases of datasets and scalability of RoPT in real-world urban traffic networks would enhance the study. While comparisons with baseline methods (e.g., CSSRNN, GETNext, NEUROMLR, Graphormer) are provided, a more in-depth comparison of computational cost and efficiency trade-offs should be included.

Some sections, particularly in the introduction and related work, contain phrases that closely resemble prior literature. Paraphrasing and rewording to highlight the unique contributions of RoPT instead of reiterating past findings will improve the originality of the text.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

Comments and Suggestions for Authors

The paper presents RoPT, a network designed to address the challenges of urban traffic management. RoPT overcomes the limitations of traditional methods, such as graph search algorithms and recurrent neural networks (RNNs), which struggle to adapt to the continuous and unpredictable changes found in real-world urban settings. By integrating Graph Convolutional Networks (GCNs) with Transformers, RoPT leverages the strengths of both to enhance route prediction accuracy. This combination of GCNs and Transformers enables the model to capture both local and global dependencies within the road network. Such a dual approach is particularly relevant in the context of urban traffic, where spatial relationships and temporal patterns are crucial. Overall, the paper is well-structured and takes a significant step forward in addressing the problem.

My Specific Remarks:

1- While the approach shows promising results, the paper could benefit from a discussion on the scalability of the model to larger datasets or more complex urban environments. This is particularly important for practical applications in real-time traffic management. Additionally, insights on how the model can be effectively deployed in real-time scenarios would greatly enhance its applicability.

2- The authors should explain how they mitigate the risk of overfitting, particularly given that multiple layers of GCNs were trained on a limited dataset. Providing insights into the model's generalization capabilities would also enhance the discussion.

3- Although the paper compares RoPT with several baseline models, including traditional algorithms like Dijkstra and advanced methods like Graphormer, it would be beneficial to compare its performance with more recent works, such as those based on Deep Heuristic Learning for Real-Time Urban Pathfinding (https://arxiv.org/html/2411.05044v1).

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 4 Report

Comments and Suggestions for Authors

This paper is entitled “RoPT: Route Planning model with Transformer” In this case, the idea and results of the paper are interesting but the following comments can be utilized to improve this paper in future.

 

Abstract

• Sentence structure is complex in some areas, making it harder to read. Example: "And, for Recurrent Neural Networks (RNN) and their variants, it is also difficult to process long sequence inputs, which results in poor reason performance and inefficient computation cost."
• Unclear phrasing in technical descriptions. Example: "With the help of the self-attention mechanism of the Transformer, the common problems of gradient disappearance and gradient explosion could be effectively solved when dealing with long-distance dependencies of sections."
• Lack of explanation on key contributions. Example: The abstract states: "features of destination position, captured from section relations, are combined with latent representations of current route for improving performance of route planning."
• Performance improvement numbers need context. Example: The abstract states: "RoPT outperforms the best methods, to the best of our knowledge, with prediction accuracy improvement of 1.49% and 1.00%, respectively."
• Stronger problem statement needed at the beginning. Example: "With the aggravation of urban traffic problems, route planning, as an important task of the traffic system, plays an increasingly critical role and becomes more urgent."

 

Experiments

• Some sentences are overly complex or ambiguous and should be rewritten for clarity. Consider all parts of the manuscript. Example: "The reason why RoPT performs better lies in that it considers not only the temporal correlations of nodes in trajectories but also spatial correlations of nodes in the road network."
• Dataset preprocessing is not explained in enough detail.
  1. How were outliers handled?
  2. What specific GPS-matching techniques were used for the Chengdu dataset?
  3. How was sparsity calculated?
  4. Were there any imbalanced data issues?
• Justify hyperparameter values with reasoning or prior research references. Example: "The number of GCN layers is set to 4."
• Improve the transition between hyperparameters and evaluation metrics. Currently, it abruptly jumps from hyperparameters to metrics.
• Clarify how baselines were adapted for fair comparisons. Example: "For the sake that this method is only designed for predicting next POI. To improve route planning performance of this method, destination information is added to the Transformer in the following experiments."
• Explain how t-SNE parameters were chosen.

 

Conclusion

• Sentence structure and phrasing need refinement for better readability. "A Route planning model, named as RoPT, is proposed for improving performance of route planning in urban transportation system."
• Avoid redundancy and repetitive phrases. Example: "Comprehensive experiments on two real-world traffic datasets, the Porto dataset and the Chengdu dataset, demonstrate that RoPT outperforms the best methods, to the best of our knowledge, with prediction accuracy improvement of 1.49% and 1.00%, respectively."
• Lacks a discussion of real-world applicability and broader impact.
  1. What are the implications of RoPT for traffic management, smart cities, or navigation apps?
  2. How does RoPT compare computationally to existing models?
  3. Are there any practical challenges in deploying RoPT?
• No mention of limitations or future research directions. Every strong conclusion should acknowledge limitations and suggest future improvements.

 

Final decision: This manuscript has interesting objectives, organization, and results. It needs major corrections before it is suitable to publish.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 4 Report

Comments and Suggestions for Authors

The comments do not addressed properly. It must be corrected again.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 3

Reviewer 4 Report

Comments and Suggestions for Authors

The author has addressed most of the comments.