MS Mamba: Spectrum Forecasting Method Based on Enhanced Mamba Architecture

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

The manuscript proposes a novel multi-scale Mamba architecture (MS-Mamba) for spectrum prediction in cognitive radio networks. The work addresses critical challenges in long-term forecasting accuracy and computational efficiency, leveraging Mamba's selective SSM capabilities. The paper is well-structured, technically sound, and presents significant innovations. Experimental validation on real-world datasets shows promising results (14.9% indoor RMSE reduction, 17% GPU memory savings). However, several areas require clarification and enhancement for publication readiness.

1. While Mamba’s application to spectrum prediction is novel, the introduction should better differentiate this work from recent Mamba adaptations in time-series forecasting (e.g., MambaTS, Cai et al., 2024). Emphasize how the dual-SSM design and multi-band pyramid specifically address spectral non-stationarity. Contrast more explicitly with Transformer variants (e.g., Pan et al., 2025) in Sect. 2.3.
2.  Clarify the number of pyramid layers (m) used and the downsampling factors. A diagram of the pyramid structure would aid understanding, and include theoretical FLOPs/parameters comparison (e.g., vs. Transformer) to supplement empirical metrics.
3. Fig. 4: Include confidence intervals for MAE/RMSE. Discuss why outdoor performance degrades more sharply (e.g., environmental dynamics). Fig. 5: Add inference latency (ms/prediction) for deployment context. Sect. 4.4: Elaborate on "adaptive weighting coefficients" in prediction heads.
4. Minor issues: Abstract: Highlight "multi-band" and "non-stationarity" more prominently.
   Sect. 1: Define "spectrum holes" for broader readability. Sect. 3.2: Justify the 10-min segmentation choice. References: Update to include Mamba literature (e.g., Gu & Dao, 2024). Typos: "in-flexibility" (Sect. 1), "con-straining" (Sect. 1).

 

 

 

 

 

 

 

 

 

 

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

This manuscript presents an enhanced approach to achieving a balance between high predictive accuracy and low computational complexity in the domain of spectrum prediction within cognitive radio systems. This research offers a solution that combines accuracy with efficiency for dynamic spectrum management, thereby enhancing spectrum utilization and reducing system energy consumption, which contributes to the advancement and practical deployment of cognitive radio technology. The manuscript, however, requires the following improvements:

1. Formatting Issues

(1) To maintain academic rigor and clarity, the first occurrence of the abbreviation “SSM” in the abstract should be accompanied by its full English expansion.

(2) The terminology used for referencing figures needs to be standardized. For instance, “Figure 3. Schematic diagram of the MS-Mamba model” should be revised to “Fig. 3” to maintain consistency with other figure references. Similarly, “Figs. 4” should be changed to “Fig. 4”.

(3) The placement of Figures 4 and 5 should be adjusted to ensure proper alignment with the corresponding textual content. Additionally, the boundaries of the charts must be strictly confined within the main text area to avoid overflow.

2. Citation Format and Literature Update

The reference formatting should be standardized to align with the submission guidelines of the journal. Moreover, considering the absence of recent developments in terahertz communication from 2023 to 2025, it is recommended to incorporate relevant literature on photonic-assisted millimeter-wave receivers to strengthen the contextual foundation of the research.

3. Enhancement of Computational Efficiency Demonstration

To more objectively illustrate the computational efficiency of the proposed method, it is recommended to include a comparative table analyzing model parameter complexity. This addition would provide quantitative evidence supporting the reduction in computational cost and reinforce the technical novelty of the proposed model architecture.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

Comments and Suggestions for Authors

1. The paper claims to be the first to apply Mamba to spectrum prediction. Does research on spectrum prediction based on SSM exist? No literature review on this aspect has been found, and it is recommended to supplement it.

 

2. In the introduction section, the literature review of the Mamba algorithm is not detailed. Why the Mamba algorithm is used and its advantages are not clearly presented in the paper.

 

3. In Section 2.1, the AR model is elaborated on without any references, which is highly unreasonable, and corresponding references should be included.

 

4. In the second section (Related Work), the advantages and disadvantages of the three algorithms should be analyzed and summarized; meanwhile, the problems to be solved by the algorithm in this paper should be highlighted.

 

5. In Section 3.2, this paper uses the Electrosense dataset but fails to clarify the data preprocessing steps, such as how to handle outliers and missing values. Furthermore, regarding the selection of the frequency band, what are the reasons stated for choosing 610-630 MHz? What is the service type of this frequency band? Are its occupancy characteristics representative?

 

6. In Section 4.3 "Dual-selective SSM Module", what are the specific parameters of the dynamic convolution? What is the value-taking rule of G? How do the parameters change dynamically? It is suggested to supplement the configuration details of the dynamic convolution.

 

7. In the simulation section of the article, what are the parameters of the comparative models? Have the comparative models undergone tuning design? Are there any comparisons with other SSM models?

 

8. The article mentions that "GPU memory is reduced by 17%", but it does not specify whether this metric is for the training phase or the inference phase. The memory consumption mechanisms of the training and inference phases are significantly different, and failure to distinguish between the scenarios will lead to distortions in the judgment of the actual deployment efficiency of the model. In addition, the efficiency variation curves under different batch sizes are not provided, making it impossible to reflect the adaptability of the model under different computational resource constraints.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 4 Report

Comments and Suggestions for Authors

The authors propose a multi-scale spectrum forecasting framework that aims to tackle challenges in long-term, multi-band spectrum prediction, which is critical for dynamic spectrum access in cognitive radio networks. The idea is interesting, and the topic is timely. Overall, the paper is well written and easy to follow.
The good results obtained, together with a reduction in computational cost, are a good contribution.

I have only a few concerns:

1) Table 1, which reports the parameters, could be extended with additional parameters. Furthermore, a column containing the parameter symbols could be added to improve the correspondence between the table and the notations used in the paper (e.g., \alpha as the learning rate).

2) In the introduction, it would be useful to emphasize more clearly why dynamic spectrum sharing is important. For example, you could refer to the machine-type communications (mMTC) scenario, where the huge number of users forces the system to switch from static to dynamic spectrum access. To support this point, you can consider, as an example, how DSA is adopted in the uplink for mMTC ([Ref1]). 

[Ref1] "Dynamic Uplink Resource Dimensioning for Massive MTC in 5G Networks Based on SCMA," European Wireless 2019; 25th European Wireless Conference, Aarhus, Denmark, 2019, pp. 1-6.

3) Ensure that acronyms are defined consistently and in only one format throughout the paper.

4) Consider adding a description of the paper’s structure at the end of the introduction.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

The authors have addressed my issues. This paper can be accepted now.

Reviewer 3 Report

Comments and Suggestions for Authors

I have no further comments. 

Reviewer 4 Report

Comments and Suggestions for Authors

The authors have addressed all concerns. It is ready for publication.