Review Reports - Intelligent Routing Optimization via GCN-Transformer Hybrid Encoder and Reinforcement Learning in Space–Air–Ground Integrated Networks

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

This paper addresses the challenging problem of intelligent routing optimization in the Space-Air-Ground Integrated Network (SAGIN). To tackle the core issues of high dynamics and strong heterogeneity in SAGIN, the authors propose a GCN-T-PPO routing algorithm that integrates a GCN-Transformer hybrid encoder with Proximal Policy Optimization (PPO) reinforcement learning. Simulation results demonstrate that the proposed scheme outperforms traditional protocols and existing intelligent routing methods in key performance metrics, including latency, packet loss rate, and QoS satisfaction rate. However, several aspects of the paper require improvement, as detailed below:

1): The abstract states that traditional and intelligent routing methods suffer from slow convergence and response, and identifies adaptive routing as the paper's goal. However, it fails to elucidate how the proposed routing algorithm achieves adaptivity. Additionally, the mechanism for predicting future network states and the specific meaning of spatiotemporal attention remain unexplained, which may lead to confusion.

2): While the paper's main contribution lies in addressing the heterogeneity and dynamism of SAGIN, the established model appears oversimplified and does not sufficiently capture these critical characteristics. The problem formulation should be strengthened to better reflect the complex nature of integrated air-space-ground networks.

3): A neural network architecture diagram is essential to illustrate how graph neural networks and Transformers are combined to extract spatiotemporal features.

4): Although the paper mentions good scalability of the GCN-T-PPO algorithm in large-scale networks, it lacks a clear analysis of real-time performance in practical deployment scenarios. It is recommended to supplement with inference latency data under different network scales to demonstrate whether the algorithm meets real-time service requirements (e.g., video streaming with latency < 100 ms). A comparative analysis of computational efficiency between the GCN-Transformer hybrid encoder and alternative approaches like Graph-Mamba would further highlight the advantages of the proposed architecture.

5): The paper incorporates dynamic weight adjustment for service priorities in the optimization objective but does not elaborate on the specific rules and adaptive mechanisms governing weight allocation. The authors should clarify how weights are adjusted in real-time according to network states (e.g., load variations, link quality fluctuations) and service types. Exploring reinforcement learning for adaptive weight parameter optimization could further enhance adaptability to differentiated QoS requirements.

6): Experimental Rigor and Comprehensive Validation:

In the ablation study, the "W/o PSO" variant employs manual parameter tuning. The specific tuning process (including parameter ranges and optimization objectives) and the resulting optimal parameter combinations should be detailed to ensure experimental fairness.
For the bursty traffic experiment using Pareto distribution with shape parameter α = 1.5, it is recommended to supplement results with different α values (e.g., 1.2, 1.8) to more comprehensively validate the algorithm's adaptability to varying burst intensities.
Comprehensive Comparison with Related Work: While the paper compares with various baseline algorithms, it would benefit from including performance comparisons with other representative GCN-Transformer hybrid models recently proposed for SAGIN routing (such as those referenced in [27-29]). This would better clarify the unique contributions of the proposed algorithm in terms of model architecture and optimization strategy, thereby strengthening the innovation claim.
The training process of the reinforcement learning algorithm and corresponding convergence curves should be provided to verify the effectiveness and stability of the proposed approach.

Comments for author File: Comments.pdf

Author Response

Comments 1: The abstract states that traditional and intelligent routing methods suffer from slow convergence and response, and identifies adaptive routing as the paper's goal. However, it fails to elucidate how the proposed routing algorithm achieves adaptivity. Additionally, the mechanism for predicting future network states and the specific meaning of spatiotemporal attention remain unexplained, which may lead to confusion.

Response 1: First, thank you very much for your careful review of this paper and for your valuable suggestions. We have provided a relatively detailed explanation in the abstract regarding the implementation of the proposed routing algorithm's adaptability, the mechanism for predicting future network states, and the specific meaning of the spatio-temporal attention mechanism. We hope this meets your requirements.

Comments 2: While the paper's main contribution lies in addressing the heterogeneity and dynamism of SAGIN, the established model appears oversimplified and does not sufficiently capture these critical characteristics. The problem formulation should be strengthened to better reflect the complex nature of integrated air-space-ground networks.

Response 2: First, thank you very much for your careful review of this paper and for your valuable suggestions. In response to your comment that the model description was overly simplified, we have added a more detailed explanation of the model at the end of Section 1.2 to more accurately characterize the heterogeneity and dynamics of SAGIN. We hope this meets your requirements.

Comments 3: A neural network architecture diagram is essential to illustrate how graph neural networks and Transformers are combined to extract spatiotemporal features.

Response 3: First, we sincerely appreciate your thorough review of this paper and the valuable feedback you provided. We fully agree with your suggestion to include a neural network architecture diagram. Due to the first author's health issues, the rest of the team has made every effort to redraw the diagram. Figure 2 in Section 3 has been replaced with the new diagram, and we hope this meets your requirements.

Comments 4: Although the paper mentions good scalability of the GCN-T-PPO algorithm in large-scale networks, it lacks a clear analysis of real-time performance in practical deployment scenarios. It is recommended to supplement with inference latency data under different network scales to demonstrate whether the algorithm meets real-time service requirements (e.g., video streaming with latency < 100 ms). A comparative analysis of computational efficiency between the GCN-Transformer hybrid encoder and alternative approaches like Graph-Mamba would further highlight the advantages of the proposed architecture.

Response 4: First, we sincerely appreciate your thorough review of this paper and the valuable feedback you have provided. Regarding your concern about the lack of real-time performance analysis in practical deployment scenarios, due to limitations in experimental conditions, funding, and the first author's health, we have made every effort to analyze the relationship between computational overhead, algorithm convergence, and networks of varying node scales in Figure 6, Table 6, and Table 7 of Section 4.5. This analysis demonstrates that our proposed algorithm exhibits superior real-time performance compared to the comparison algorithms, offering certain advantages, though it may not fully meet your expectations. Therefore, we have incorporated your suggestion into our future research plans in Section 5 (Conclusion). Once the first author has recovered, we will convey your valuable feedback to him and integrate it into our subsequent research.

Comments 5: The paper incorporates dynamic weight adjustment for service priorities in the optimization objective but does not elaborate on the specific rules and adaptive mechanisms governing weight allocation. The authors should clarify how weights are adjusted in real-time according to network states (e.g., load variations, link quality fluctuations) and service types. Exploring reinforcement learning for adaptive weight parameter optimization could further enhance adaptability to differentiated QoS requirements.

Response 5: First, thank you very much for your careful review of this paper and for your valuable feedback. Regarding your point about the lack of detailed rules for weight distribution and the adaptive mechanism, we have made every effort to address this in Section 2.2. This section now includes rules for how weights are dynamically adjusted in real-time based on network conditions (such as load fluctuations and link quality variations) and service types. Combined with the subsequent experimental results analysis, this demonstrates that our proposed algorithm exhibits superior real-time performance compared to the baseline algorithms, offering distinct advantages.

Comments 6: Experimental Rigor and Comprehensive Validation:

(1) In the ablation study, the "W/o PSO" variant employs manual parameter tuning. The specific tuning process (including parameter ranges and optimization objectives) and the resulting optimal parameter combinations should be detailed to ensure experimental fairness.

(2) For the bursty traffic experiment using Pareto distribution with shape parameter α = 1.5, it is recommended to supplement results with different α values (e.g., 1.2, 1.8) to more comprehensively validate the algorithm's adaptability to varying burst intensities.

(3) Comprehensive Comparison with Related Work: While the paper compares with various baseline algorithms, it would benefit from including performance comparisons with other representative GCN-Transformer hybrid models recently proposed for SAGIN routing (such as those referenced in [27-29]). This would better clarify the unique contributions of the proposed algorithm in terms of model architecture and optimization strategy, thereby strengthening the innovation claim.

(4) The training process of the reinforcement learning algorithm and corresponding convergence curves should be provided to verify the effectiveness and stability of the proposed approach.

Response 6:

(1) Thank you very much for your careful review of this paper and for your valuable suggestions. We fully acknowledge this feedback and have included a description of the manual parameter tuning process for the “W/o PSO” variant in Section 4.4—Ablation Study, along with the optimal parameter combination.

(2) First, we sincerely appreciate your thorough review of this paper and the valuable suggestions you have provided. We fully agree with your recommendation to supplement the experimental results with burst traffic at different Pareto α values to more comprehensively validate the algorithm's effectiveness in adapting to varying burst intensities. Therefore, we have configured the experiment for Scenario 2 in Section 4.3 as follows: the total network load is fixed at 70%, and burst traffic is generated using the aforementioned Pareto traffic model. The X-axis represents burst intensity (higher values indicate stronger bursts), while the Y-axis shows the real-time video service's QoS fulfillment rate (i.e., the proportion of packets with latency < 100 milliseconds). The obtained results clearly demonstrate the advantages of our proposed model over other models.

(3) First, we sincerely appreciate your thorough review of this paper and the valuable suggestions you have provided. We fully acknowledge your recommendation to supplement the performance comparisons with other representative GCN-Transformer hybrid models recently proposed for SAGIN routing (such as those mentioned in [27-29]). However, the models referenced in [27-29] differ significantly from the routing domain investigated in this paper. The proposed method represents a novel intelligent routing algorithm for integrated air-ground-space communication networks, developed by reviewing baseline algorithms compared herein and analyzing relevant literature from the past three years. Preliminary research indicates this field is still in its infancy. This algorithm represents a relatively cutting-edge approach within the field of intelligent routing for integrated air-ground-space communication networks. Experimental results fully validate the advantages of the proposed algorithm. Due to the first author's health condition, more in-depth comparative experiments could not be conducted promptly. We have included these plans in Section 5 (Future Work). Once the author recovers, we will promptly relay your suggestions to him and encourage him to pursue more in-depth research in subsequent studies.

(4) First, we sincerely appreciate your thorough review of this paper and the valuable suggestions you have provided. We fully agree with your recommendation that we should present the training process of the reinforcement learning algorithm and the corresponding convergence curves to validate the effectiveness and stability of the proposed method. However, this paper has already conducted experiments on the relationship between algorithm convergence/adaptation time and the total number of network nodes (N). The experimental results sufficiently demonstrate the advantages of the proposed algorithm. Given the first author's health-related circumstances, conducting more in-depth experiments promptly is not feasible. We have therefore included this topic in Section 5 (Future Work). Once the author recovers, we will promptly relay your suggestions to him and encourage him to pursue more in-depth research in subsequent studies.

4. The English could be improved to more clearly express the research

Point: The English could be improved to more clearly express the research

Response: First, thank you very much for your careful review of this paper and for your valuable suggestions. We fully agree with your recommendation to improve the English expression in the article. Due to the first author's health condition, the rest of the team has made every effort to read the entire text and revise any non-standard expressions, hoping to meet your requirements.

5. Additional clarifications

Due to the nature of the first author's profession, he unfortunately sustained an injury during training and has just undergone surgery. As a result, he is unable to promptly implement the valuable revisions suggested by you. Our team sincerely apologizes for this delay. Once he has recovered, we will incorporate your valuable feedback into his future research efforts. Once again, we sincerely apologize. The rest of the team has made every effort to revise the manuscript. Approaching graduation,we earnestly hope you will accept this paper and extend your support to this unfortunate author, enabling him to meet his graduation requirements.

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

This paper focuses on the intelligent routing optimization of the 6G core architecture SAGIN, with precise topic selection, clear technical roadmap, and sufficient experimental verification. It is a high-quality academic research.
But before receiving, I think the following content needs to be modified:
1. The algorithm performance is highly dependent on the prediction accuracy of GCN Transformer (current MSE=0.04), but in practical scenarios, there may be "unpredictable sudden anomalies" (such as link interruptions caused by extreme weather, temporary satellite failures).
2. The paper only sets two types of QoS requirements: "real-time video" and "best effort service", while SAGIN actually carries richer services (such as low power consumption in the Internet of Things, high reliability in the Internet of Vehicles, and satellite remote sensing big data transmission).
3. Reference 15 is incomplete.
I think the following references can be added to enrich the preliminary research: Guard band protection for coexistence of 5G base stations and satellite earth stations. Authors：Shuzhi Liu, Yiqiao Wei, Seung-Hoon Hwang
The core value of this paper lies in "using precise technology to integrate and solve specific industry pain points". Its concept of "spatiotemporal prediction+active decision-making", mixed model design logic, and comprehensive experimental verification paradigm have important reference significance for research in fields such as 6G networks, satellite communication, and intelligent routing. Suggest accepting after modification.

Author Response

Comments 1: The algorithm performance is highly dependent on the prediction accuracy of GCN Transformer (current MSE=0.04), but in practical scenarios, there may be "unpredictable sudden anomalies" (such as link interruptions caused by extreme weather, temporary satellite failures).

Response 1: Thank you for pointing out this critical issue. We fully agree that unpredictable anomalies are a key challenge in real-world SAGIN deployment. In the revised version, we have:

Added a new paragraph in Page 6 Section 2.3 (Challenge Modeling), Paragraph 4 to explicitly discuss the limitations of current prediction models in handling such anomalies；

Added a new item in Page 21,Section 5 (Conclusion and Future Work), Paragraph 2 to propose the integration of robust RL and real-time anomaly detection as future research directions.

Comments 2: The paper only sets two types of QoS requirements: "real-time video" and "best effort service", while SAGIN actually carries richer services (such as low power consumption in the Internet of Things, high reliability in the Internet of Vehicles, and satellite remote sensing big data transmission).

Response 2: Agree. We appreciate your suggestion to enrich the QoS modeling. However, due to the nature of the first author's profession, he unfortunately sustained an injury during training and has just undergone surgery. As a result, he is unable to promptly implement the valuable revisions suggested by you. Our team sincerely apologizes for this delay. Once he has recovered, we will incorporate your valuable feedback into his future research efforts. Once again, we sincerely apologize.

Comments 3: Reference 15 is incomplete.
I think the following references can be added to enrich the preliminary research: Guard band protection for coexistence of 5G base stations and satellite earth stations. Authors：Shuzhi Liu, Yiqiao Wei, Seung-Hoon Hwang.

Response 3: Thank you for your careful check. We have:

Completed Reference 15 with full bibliographic information；

Added the suggested paper by Liu et al. on guard band protection for 5G-satellite coexistence；

Cited this paper in Page 2, Section 1.3(Strong Heterogeneity), Paragraph 1 to support the discussion on spectrum coordination in heterogeneous networks.

4. Additional clarifications

Author Response File: Author Response.pdf

Reviewer 3 Report

Comments and Suggestions for Authors

In this paper, the authors present a routing framework for Space–Air–Ground Integrated Networks (SAGIN) that combines a hybrid GCN–Transformer encoder for high-precision spatiotemporal prediction with a PPO agent enhanced through a spatiotemporal attention mechanism. The proposal is evaluated through extensive experiments based on realistic satellite mobility extracted from CelesTrak TLE data, UAV mobility traces, and CAIDA-derived traffic models. The integration of GNNs with Transformer-based structures shows familiarity with modern dynamic graph learning, and the use of Particle Swarm Optimization for hyperparameter tuning is a reasonable choice given the dimensionality and interdependence of model parameters. The organization and writing are generally sufficient.

There are some enhancements that can improve the quality of the paper.

1) Some parts of the paper rely too much on a bullet-point style, especially in the introduction, and could be written in a more discursive manner to improve readability.

2) There are several considerations regarding the decision to adopt a fully centralized approach. Both training and execution rely on a G-CCC that collects telemetry from all nodes and issues routing instructions. This design introduces a single point of failure and may face substantial scalability challenges in large or geographically distributed SAGINs. Adding some discussion of these aspects would strengthen the paper.

3) Throughout the paper, including the abstract, there is the presence of long dashes. It would be better to remove them anche change with comma.

4) The clarity of the mathematical expressions could be improved. For example, sets are referenced without using a consistent notation such as the math style

5) The proposed architecture integrates many different components, resulting in a pipeline that is computationally heavy and operationally intricate. Although the authors justify each component individually, the paper does not clearly demonstrate the benefit achieved relative to the added computational complexity. It would be useful to include metrics that quantify the advantages of the proposed system by relating performance gains to the associated costs. For example, the weighted complexity metric adopted in 10.1109/TVT.2022.3216028 could serve as a reference. Incorporating such considerations would help readers better understand the overall impact of the proposal.

6) The paper does not measure or discuss this overhead, which is a crucial factor for the viability of centralized routing in satellite networks, where bandwidth is limited and downlink/uplink times are non-negligible. Even if a precise quantification is not provided, offering some discussion of this aspect would strengthen the work.

7) The quality of the figures could be improved, and the punctuation of equations could also be refined.

8) It is not clear the role of Source/Notes in Table 2, given that it is almost empty.

Author Response

Comments 1: Some parts of the paper rely too much on a bullet-point style, especially in the introduction, and could be written in a more discursive manner to improve readability.

Response 1: Thank you for pointing out this critical issue. In the revised version, we have:

Thoroughly revised the overly bullet-pointed style in sections 1.2, 1.3, 1.4, and 1.5 of the paper's introduction, adopting a more discursive writing approach to enhance readability. For the remaining bullet-pointed sections, we believe this structure better maintains logical coherence and clarity, and thus no changes were made.

Comments 2: There are several considerations regarding the decision to adopt a fully centralized approach. Both training and execution rely on a G-CCC that collects telemetry from all nodes and issues routing instructions. This design introduces a single point of failure and may face substantial scalability challenges in large or geographically distributed SAGINs. Adding some discussion of these aspects would strengthen the paper.

Response 2: Agree. We greatly appreciate your suggestion regarding the potential risks of “single point of failure” and “scalability” associated with the centralized architecture proposed in this paper. However, due to the nature of the first author's profession, he unfortunately sustained an injury during training and has just undergone surgery. Consequently, he is temporarily unable to promptly implement the valuable revision suggestions you have provided. Our team sincerely apologizes for this delay. We have incorporated the relevant discussion into Section 5, “CONCLUSION-Future Work.” Once he has recovered, we will integrate your valuable insights into his future research endeavors. We extend our deepest apologies once again.

Comments 3: Throughout the paper, including the abstract, there is the presence of long dashes. It would be better to remove them anche change with comma.

Response 3: Thank you for your careful check. We have:

After carefully reading and reviewing the entire text, we identified a total of seven similar errors (located in the abstract, introduction, conclusion, etc.) and made logical corrections to them.

Comments 4: The clarity of the mathematical expressions could be improved. For example, sets are referenced without using a consistent notation such as the math style.

Response 4: First of all, thank you very much for your thorough review and valuable suggestions. We have carefully proofread and strictly revised each formula according to the official paper template to ensure compliance with journal requirements.

Comments 5: The proposed architecture integrates many different components, resulting in a pipeline that is computationally heavy and operationally intricate. Although the authors justify each component individually, the paper does not clearly demonstrate the benefit achieved relative to the added computational complexity. It would be useful to include metrics that quantify the advantages of the proposed system by relating performance gains to the associated costs. For example, the weighted complexity metric adopted in 10.1109/TVT.2022.3216028 could serve as a reference. Incorporating such considerations would help readers better understand the overall impact of the proposal.

Response 5: First, we sincerely appreciate your thorough review and valuable suggestions regarding the quantification of the relationship between increased computational complexity and performance gains. However, due to the nature of the first author's profession, he unfortunately sustained an injury during training and has just undergone surgery. Consequently, he is temporarily unable to promptly implement your valuable revision suggestions. Our team sincerely apologizes for this delay. The relevant discussion has been incorporated into Chapter 5, “Conclusions and Future Work.” Once he has recovered, we will integrate your valuable insights into his future research endeavors. We extend our deepest apologies once again.

Comments 6: The paper does not measure or discuss this overhead, which is a crucial factor for the viability of centralized routing in satellite networks, where bandwidth is limited and downlink/uplink times are non-negligible. Even if a precise quantification is not provided, offering some discussion of this aspect would strengthen the work.

Response 6: First, we sincerely appreciate your thorough review and valuable feedback regarding routing overhead. However, we have already introduced experiments and discussions on computational overhead versus total network nodes in Section 4.5 Scalability Test, which we believe provides a compelling basis—though it may not fully meet your expectations. Unfortunately, due to the nature of the first author's profession, he sustained an injury during training and recently underwent surgery. Consequently, he is temporarily unable to promptly implement your valuable revision suggestions. Our team sincerely apologizes for this delay. The relevant discussion has been incorporated into Chapter 5, “Conclusions and Future Work.” Once he has recovered, we will integrate your valuable insights into his future research endeavors. We extend our deepest apologies once again.

Comments 7: The quality of the figures could be improved, and the punctuation of equations could also be refined.

Response 7: First, thank you very much for your thorough review and valuable suggestions regarding image clarity and formula punctuation standards. We have made every effort to enhance the clarity of all images and replaced them accordingly. Subsequently, we added punctuation marks at the end of each formula in accordance with the journal paper template, hoping to meet your requirements.

Comments 8: It is not clear the role of Source/Notes in Table 2, given that it is almost empty.

Response 8: First, thank you very much for your careful review and valuable feedback regarding the source/annotation section of Table 2 in the paper. Our original intention was to add the corresponding data and module structure sources, but since the paper already contains relevant descriptions of these, we have removed them here.

4. Additional clarifications

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

We appreciate the authors' efforts to revise and respond to the initial review comments, especially given the special circumstance where the first author was injured and unable to fully participate in the revision. However, key issues such as ambiguous definitions of graph neural networks (GNNs) and the lack of direct verification of the training convergence of spatiotemporal-coupled neural networks have not been substantially resolved. Some experimental supplements are limited, and the overall academic rigor and persuasiveness of the results still need improvement. The manuscript does not meet the journal's acceptance criteria for the time being and requires further revision.

Major Remaining Concerns:

（1）：Response to Comment 3 is insufficient. Merely providing an architecture diagram (Figure 2) is a first step but does not constitute a clear definition. The manuscript must explicitly detail:

The precise formulation of the Graph Convolutional Network (GCN) used. What is the convolutional operator (e.g., spectral or spatial)? What is the exact message-passing or aggregation function?
How is the network topology graph constructed for the GCN? What are the nodes and edges, and how are their features defined (e.g., node features: buffer status, location? edge features: link delay, capacity?)?

（2）：Response to Comment 6(4) is critically deficient and unacceptable for publication. The statement that "experiments on algorithm convergence/adaptation time... sufficiently demonstrate the advantages" completely misses the point. The training process details and convergence curves of the Reinforcement Learning (RL) algorithm itself are required, not just post-training adaptation time. For a novel hybrid model combining GCN, Transformer, and PPO, demonstrating that the training process is stable and converges to a good policy is paramount. The authors need provide:

Plots of the training reward/return vs. training episodes/iterations for their proposed model and key baselines.
Discussion of training hyperparameters for the RL component (e.g., learning rates, discount factor, PPO clipping parameter).
Analysis of training stability and sample efficiency. Without this, the claim of a working, trainable model is not scientifically verifiable. The suggestion to move this to "future work" is not viable for the current manuscript.

(3) While the adoption of PSO for hyperparameter optimization is a sensible step, the manuscript must address two key methodological gaps to substantiate its effectiveness. First, given the high-dimensional and mixed-type search space (e.g., layers, heads, learning rate), details on the PSO configuration (swarm size, iteration budget, convergence behavior) and the management of its substantial computational cost are necessary to confirm a thorough and efficient search was conducted. More critically, the authors must demonstrate the robustness and generalizability of the single hyperparameter set obtained. It is highly plausible that architecture-sensitive parameters (like the number of GCN layers) optimized for one specific network scale may not transfer optimally to networks of different sizes. Without validating the same parameter set across the varying scales used in the experiments or providing a sensitivity analysis, the performance comparisons across scenarios remain confounded. Proving the parameter set's robustness is essential to isolate and verify the inherent advantages of the proposed model architecture itself.

Comments for author File: Comments.pdf

Author Response

Comments 1: Response to Comment 3 is insufficient. Merely providing an architecture diagram (Figure 2) is a first step but does not constitute a clear definition. The manuscript must explicitly detail:

The precise formulation of the Graph Convolutional Network (GCN) used. What is the convolutional operator (e.g., spectral or spatial)? What is the exact message-passing or aggregation function?How is the network topology graph constructed for the GCN? What are the nodes and edges, and how are their features defined (e.g., node features: buffer status, location? edge features: link delay, capacity?)?

Response 1: First, we sincerely appreciate your understanding of our current predicament and your valuable feedback on this paper. Regarding your comment that our description of the GCN network was insufficiently precise, we have made every effort to address this in the following sections: the beginning of Section 2.1. Network Model, Section 3.1. Network State Prediction Model (GCN+Transformer+PSO). These additions include: Detailed definitions of nodes, edges, and their properties within the GCN network; Clarification that the convolutional operator used is the spatial convolution operator; Specification of the message passing process and aggregation function; Explanation of the GCN network topology construction process. We hope you understand our current constraints and that these revisions meet your expectations.

Comments 2: Response to Comment 6(4) is critically deficient and unacceptable for publication. The statement that "experiments on algorithm convergence/adaptation time... sufficiently demonstrate the advantages" completely misses the point. The training process details and convergence curves of the Reinforcement Learning (RL) algorithm itself are required, not just post-training adaptation time. For a novel hybrid model combining GCN, Transformer, and PPO, demonstrating that the training process is stable and converges to a good policy is paramount. The authors need provide:

Plots of the training reward/return vs. training episodes/iterations for their proposed model and key baselines.

Discussion of training hyperparameters for the RL component (e.g., learning rates, discount factor, PPO clipping parameter).

Analysis of training stability and sample efficiency. Without this, the claim of a working, trainable model is not scientifically verifiable. The suggestion to move this to "future work" is not viable for the current manuscript.

Response 2: First, we sincerely appreciate your understanding of our current constraints and your valuable feedback on this paper. Regarding your suggestions: we have provided charts illustrating the relationship between training rewards/payoffs and training episodes/iterations for the model and key baselines; discussed training hyperparameters for reinforcement learning components (e.g., learning rate, discount factor, PPO clipping parameters); and addressed issues concerning training stability and sample efficiency analysis. We have made every effort to address these points by adding detailed explanations at the end of Section 4.5. Scalability Test. This includes training performance of the GCN-T-PPO algorithm under different learning rates, as well as comparative training performance of various algorithms under identical experimental conditions (including training results plots and analysis). Results indicate that the proposed algorithm achieves optimal rewards at a learning rate of 0.001, with rewards stabilizing without significant fluctuations after 250 epochs. Comparisons across algorithms show rewards stabilizing without noticeable oscillations after 200 epochs, with the proposed algorithm demonstrating certain advantages. We hope you understand our current constraints and that these revisions meet your requirements.

Comments 3: Response to Comment 6(4) is critically deficient and unacceptable for publication. The statement that "experiments on algorithm convergence/adaptation time... sufficiently demonstrate the advantages" completely misses the point. The training process details and convergence curves of the Reinforcement Learning (RL) algorithm itself are required, not just post-training adaptation time. For a novel hybrid model combining GCN, Transformer, and PPO, demonstrating that the training process is stable and converges to a good policy is paramount. The authors need provide:

Plots of the training reward/return vs. training episodes/iterations for their proposed model and key baselines.

Discussion of training hyperparameters for the RL component (e.g., learning rates, discount factor, PPO clipping parameter).

Response 3: First, we sincerely appreciate your understanding of our current predicament and your valuable feedback on this paper. Regarding your request for details on the PSO configuration (population size, iteration budget, convergence behavior), management of its high computational cost, and the robustness and generalizability of the obtained single hyperparameter set, we have made every effort to address these points in Section 3.1. Network State Prediction Model (GCN+Transformer+PSO). Subsequently, robust testing experiments using different hyperparameter optimization methods across varying network scales were added at the end of Section 4.6. Robustness Test. Results are summarized in Table 8, demonstrating that PSO hyperparameter optimization exhibits superior robustness and generalization compared to manual and random hyperparameter optimization. We hope you understand our current constraints and that these revisions meet your requirements.

4. The English could be improved to more clearly express the research

Point: The English could be improved to more clearly express the research

5. Additional clarifications

Author Response File: Author Response.pdf

Reviewer 3 Report

Comments and Suggestions for Authors

Dear authors,
I am sorry for the issue experienced by the first author, and I understand that this made it difficult to work on the revisions. However, since it is important to bring the quality of the paper to an adequate level, I propose conducting another round of review to address the remaining concerns and allow some additional time to complete the work.

The proposed architecture integrates many different components, resulting in a pipeline that is computationally heavy and operationally intricate. Although the authors justify each component individually, the paper does not clearly demonstrate the benefit achieved relative to the added computational complexity. It would be useful to include metrics that quantify the advantages of the proposed system by relating performance gains to the associated costs. For example, the weighted complexity metric adopted in 10.1109/TVT.2022.3216028 could serve as a reference. Incorporating such considerations would help readers better understand the overall impact of the proposal.
The paper does not measure or discuss this overhead, which is a crucial factor for the viability of centralized routing in satellite networks, where bandwidth is limited and downlink/uplink times are non-negligible. Even if a precise quantification is not provided, offering some discussion of this aspect would strengthen the work.

Author Response

Comments 1: The proposed architecture integrates many different components, resulting in a pipeline that is computationally heavy and operationally intricate. Although the authors justify each component individually, the paper does not clearly demonstrate the benefit achieved relative to the added computational complexity. It would be useful to include metrics that quantify the advantages of the proposed system by relating performance gains to the associated costs. For example, the weighted complexity metric adopted in 10.1109/TVT.2022.3216028 could serve as a reference. Incorporating such considerations would help readers better understand the overall impact of the proposal.

Response 1: First, we sincerely appreciate your thorough review of this paper and the valuable feedback you have provided. Addressing the issue of “quantifying computational complexity versus performance gains,” we have added Section 4.7, “Complexity-Performance Tradeoff Optimization,” to the paper and included corresponding experiments. The experimental results demonstrate that the method proposed in this paper achieves more significant QoS performance improvements compared to other algorithms, albeit at a slightly higher computational complexity. This validates the rationality of the “spatiotemporal prediction + attention mechanism” architecture.

Comments 2: The paper does not measure or discuss this overhead, which is a crucial factor for the viability of centralized routing in satellite networks, where bandwidth is limited and downlink/uplink times are non-negligible. Even if a precise quantification is not provided, offering some discussion of this aspect would strengthen the work.

Response 2: First, thank you very much for your careful review of this paper and for your valuable comments. Regarding the issue of “pointing out that the paper did not measure/discuss the overhead of centralized routing,” we have addressed this in Section 1.3. Core Challenges of SAGIN Routing, added an explanation of centralized control and execution architecture in Section 3.3, included an analysis of the overhead in Section 4.5. Scalability Test, and finally supplemented Section 5. CONCLUSION with a statement on measuring centralized routing overhead as a future research task.

4. Additional clarifications

Author Response File: Author Response.pdf

Round 3

Reviewer 3 Report

Comments and Suggestions for Authors

Thank you for the efforts you dedicated to improving the quality of the work. I believe that it is now sufficient. As a suggestion, you may further improve the quality of Figs. 1, 6, and 7.

Author Response

Comments 1: As a suggestion, you may further improve the quality of Figs. 1, 6, and 7.

Response 1: First, we sincerely appreciate your understanding of our current predicament and your valuable feedback on this paper. Regarding your suggestion that we further improve the quality of Figures 1, 6, and 7, we have made every effort to enhance these three images and have updated them accordingly. We hope you will understand our current difficulties and that these revisions meet your expectations.

Author Response File: Author Response.docx