MAT-WGCN: Traffic Speed Prediction Using Multi-Head Attention Mechanism and Weighted Adjacency Matrix

Round 1

Reviewer 1 Report

The topic is of interest, but the flow of the paper and the various equations are difficult to be understandable to an average reader.

What will be practical applications of this research work is not clear.

There are no units specified for axis titles in figures and some of figures are not clear at all.

Moderate English corrections are required.

Author Response

Please see the attachment. Thank you

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments for Authors

-          The English of the paper is not clear in several parts, and some parts are not clear enough to understand the authors' idea. The English should be improved and the grammatical mistakes should be corrected.

-          The introduction provides a comprehensive overview of the importance of intelligent transportation systems in modern urban traffic management and planning. The authors effectively highlight the challenges posed by increasing urban populations and vehicle numbers, necessitating accurate traffic forecasting methods. The distinction between traditional and deep learning methods is well-explained, showcasing the limitations of conventional approaches in predicting traffic flow. Additionally, the review of recent advancements in deep learning techniques demonstrates the evolving landscape of traffic speed prediction. The proposed MAT-WGCN model seems promising, with its emphasis on incorporating road length as a critical factor in spatial feature extraction. The introduction effectively outlines the use of GCN, GRU, and multi-head attention mechanism in constructing the model, showcasing the potential benefits of this approach. The authors' focus on testing the model on real-world data sets adds credibility to their claims.

 

o   Could you provide more insights into the specific challenges posed by increasing urban populations and vehicle numbers? How do these challenges impact the effectiveness of existing traffic forecasting methods?

o   In the comparison between traditional methods and deep learning methods, the introduction mentions that deep learning methods enable traffic speed prediction. Could you elaborate on how deep learning methods achieve this, and how they overcome the limitations of traditional approaches in capturing temporal dependencies?

o   While the proposed MAT-WGCN model addresses the shortcomings of conventional models by considering road length, could you explain how the weighted adjacency matrix enhances spatial feature extraction? How is the road length information incorporated into the matrix, and how does it improve the model's performance?

o   The multi-head attention mechanism is mentioned as a critical component of the MAT-WGCN model. Could you elaborate on its operation and how it aids in capturing temporal and spatial relationships in traffic speed prediction?

o   The introduction mentions real data sets used to test the MAT-WGCN model. Could you provide details about the data sets, including their size, diversity, and relevance to real-world traffic conditions? How do these data sets compare with those used in previous studies?

 

Overall, the introduction presents a clear and informative overview of the paper's objectives and the proposed MAT-WGCN model. However, additional details and explanations are needed to fully understand the model's novelty and potential contributions to the field of traffic speed prediction. Addressing the questions above would help to strengthen the introduction and provide valuable insights into the research work.

 

-          The section on "Multi-Head Attention Mechanism" provides a clear explanation of the attention mechanism and its significance in processing sequence or ensemble data. The historical background on its development, starting with machine translation models and later advancements like the Self-Attention mechanism in the Transformer model, adds context to its widespread application. The introduction of the graph attention mechanism and the subsequent development of the multi-head attention mechanism by Petar et al. to enhance model stability and robustness are also well-described. The explanation of how the multi-head attention mechanism can balance the sensitivity of features by assigning weights to temporal and spatial features is particularly insightful. The section effectively outlines how attention weights can highlight important time points or spatial regions, resulting in improved task performance and representation.The discussion of the key role of hidden state weight selection in the multi-head attention mechanism, specifically in the MAT-WGCN model, is important for understanding its implementation. The section highlights how MAT-WGCN employs the multi-head attention mechanism to assign weights to spatial and temporal features after GRU and GCN processing, ultimately leading to improved performance.

o   Could you provide more specific examples of tasks in various fields where the attention mechanism has demonstrated powerful capabilities? How do these examples relate to traffic speed prediction or other transportation-related tasks?

o   The section mentions that the graph attention mechanism was developed to enhance stability and robustness. Could you explain the potential challenges or limitations faced by traditional attention mechanisms that the graph attention mechanism addresses?

o   While the multi-head attention mechanism allows the model to focus on different temporal and spatial information, how does it ensure that these multiple heads do not lead to redundancy or conflicting weight assignments? Are there any techniques implemented to handle potential conflicts among the different heads?

o   In the context of MAT-WGCN, what specific advantages does the multi-head attention mechanism offer over single-head attention mechanisms? Are there any trade-offs or computational complexities associated with employing multiple heads?

o   The section mentions that MAT-WGCN uses the attention mechanism to assign weights to hidden states. Could you elaborate on how these weighted hidden states contribute to the overall traffic speed prediction task? How are these weighted hidden states utilized in the final prediction process?

-          A flow chart of the developed method is needed to make it more understandable.

-          Please ensure consistency in using different key terms.

-          The literature review should be enhanced by presenting a critical review, not just presenting information about who did what. The authors should prepare a table to highlight the previous contributions and research gaps in a more robust way. The recent references must be cited and explained. The references from top journals should be explored.

-          Identifying key factors should not be a part of the literature review. It should be a part of the result analysis. Literature review analyses the closely related papers to identify research gaps.

-          The result analysis should be improved based on the unique findings, interesting insights, and how these results will be useful to the practice.

-          The conclusion section can be revised considering unique findings, contributions, limitations, and future research directions.

-          Check the citations and references (one by one) if there is any missing information. Citations and references must be 100% accurate.

-          The English of the paper is not clear in several parts, and some parts are not clear enough to understand the authors' idea. The English should be improved and the grammatical mistakes should be corrected.

Author Response

Please see the attachment. Thank your

Author Response File: Author Response.pdf

Reviewer 3 Report

The work seems good and title is attractive but long, needs to lower down with reduction in words

Mention in the intro that how the proposed work is different than a) Traffic Speed Prediction Based on Heterogeneous Graph Attention Residual Time Series Convolutional Networks and b) Design and implementation of an ML and IoT based adaptive traffic-management system for smart cities 

Contributions are good way mentioned and described here. I appreciate that.

Experimentation needs more explanation and needs more comparisons for better paper.

Some recent work can also be used like "A Smart Cloud and IoVT-Based Kernel Adaptive Filtering Framework for Parking Prediction" to help, read and manage the paper in a good way.

Future work needs to be mentioned and conclusion can be little less in explanation.

Technique with analytical analysis is much more appreciated.

Author Response

Please see the attachment. Thank you

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

All the comments have been addressed, however some of the figures need refinement in terms of proper axis title and descriptions.

Average

Reviewer 2 Report

-          Check the citations and references (one by one) if there is any missing information. Citations and references must be 100% accurate.

-          The English of the paper is good in several parts, and some parts are enough to understand the authors' idea. But the English can be improved more.