Insights into Multi-Model Federated Learning: An Advanced Approach for Air Quality Index Forecasting
, Anh-Khoa Tran 2, Minh-Son Dao 2,*, Kieu-Chinh Nguyen-Ly 1, Hoang-Son Le 1, Xuan-Dao Nguyen-Thi 1, Thanh-Qui Pham 1, Van-Luong Nguyen 1 and Bach-Yen Nguyen-Thi 1
Round 1
Reviewer 1 Report
This paper is regarding the review of forecasting air quality index using a federated learning approach. This is a review paper and not an article. The paper presents good insights about the topics but lacks in the following things:
1. The referencing is not consistent. It should be in order e.g. [1-24, 45-67]. This is not the correct way just make it [1-50] e.g. On the other hand, this is too much of a big number to put all these papers in one reference. In such kind of a review paper, a better practice is to make a comparative table of the previous articles with their key points and shortcoming. Then it is easier for the reader to grasp like table 1. It's a bit late presented in the article, it should be a bit earlier.
2. Sections 1 and 2 can be merged.
3. Section 4 can also be merged with the previous and made concise. This is not a very good heading "more about FL". Or it can be made with some better wording.
4. In a review paper, the conclusion and discussion parts must be very strong and they should provide some clear answers or guidelines. Therefore, the conclusion section can be revised and more discussions can be added.
Author Response
Dear Reviewer. Thank you so much for your comments. We appreciated them. We adjusted the paper as follows:
- The referencing is not consistent. It should be in order e.g. [1-24, 45-67]. This is not the correct way just make it [1-50] e.g. On the other hand, this is too much of a big number to put all these papers in one reference. In such kind of a review paper, a better practice is to make a comparative table of the previous articles with their key points and shortcoming. Then it is easier for the reader to grasp like table 1. It's a bit late presented in the article, it should be a bit earlier.
Yes, we fixed the referencing. We also made Table 1 come sooner, it is now on the 2nd page instead of the 3rd page. Thank you
- Sections 1 and 2 can be merged.
Yes, we merged the two sections in to section 1 “Traditional approaches in AQI prediction”, please see it at line 14. Thank you
- Section 4 can also be merged with the previous and made concise. This is not a very good heading "more about FL". Or it can be made with some better wording.
We use another word for this section: “Insights into Federated Learning” at line 162. We shorten section 4 by moving the subsection Challenges and Benchmarks to new sections. Please see them at line 262 and 356. Thank you
- In a review paper, the conclusion and discussion parts must be very strong and they should provide some clear answers or guidelines. Therefore, the conclusion section can be revised and more discussions can be added.
Yes, we rewrote the conclusion. Please see it at line 513. Thank you
Author Response File: 
Author Response.pdf
Reviewer 2 Report
This work finds there is a lack of survey work on the broader perspective of AQI predictions to cover possible future research in this area. The work concludes that multi-model federated learning is the most effective technique that could enhance the AQI prediction result.
1- The motivation is ill-suited as highlighted in Line 61 for two reasons: 1) is FL the right use-case for AQI forecasting, is the data collected private and can not be shared; 2) this paper does not cover any paper using FL for AQI forecasting?
2- Both Tables 1 and 2 are hard to interpret and read
3- Only 4 works related to the use of FL in AQI - hence it does suffice for the survey purpose, moreover, the paper did not give us a clear distinction between different strategies or qualitative comparisons among those papers. These works are summarised in Table 2 and are never discussed in detail in the text. The paper focused on the comparison of FL architectures instead which has been widely considered by many previous surveys as stated by the paper.
4- In table 2, what are evaluation meters? is it metrics instead?
5- In Table 2, the works presented, How the energy consumption reduced? These works are questionable in terms of their claim as their energy consumption numbers do not seem realistic and are not based on accurate energy models or real measurements. So it is not clear whether the conclusions made are appropriate.
6- The heterogeneity challenge in line 231, Lacks citations of works that studied the heterogeneity impact such as https://dl.acm.org/doi/abs/10.1145/3517207.3526969 or https://dl.acm.org/doi/abs/10.1145/3442381.3449851 and another work that leveraged adaptive model quantization to address it such as https://dl.acm.org/doi/abs/10.1145/3437984.3458839
7- In future works, Not clear what is the motivation for multi-model esp. for the AQI use-case. The authors cite works that perform Multimodal learning in AQI prediction use-case but this has no relevance to the proposed multi-model FL approaches. It seems the authors also fail to make the distinction clear between multi-modal, multi-model, and multi-task learning.
8- The work concludes that multi-model federated learning is the best approach without giving any solid observations or supporting evidence. It would be beneficial if some experimental evaluations are used to support the claims.
9- Finally, the authors make an over-claim by stating that they surveyed 90 papers in the scope of AQI predictions which is not the case. Having 90 citations in the work does not mean you have surveyed them all.
Author Response
Dear Reviewer. Thank you so much for your comments. We appreciated them. We adjusted the paper as follows:
1- The motivation is ill-suited as highlighted in Line 61 for two reasons: 1) is FL the right use-case for AQI forecasting, is the data collected private and can not be shared; 2) this paper does not cover any paper using FL for AQI forecasting?
We removed the ill-suited motivation.
2- Both Tables 1 and 2 are hard to interpret and read
We explain the columns of Table 1 at line 23-27. Thank you
3- Only 4 works related to the use of FL in AQI - hence it does suffice for the survey purpose, moreover, the paper did not give us a clear distinction between different strategies or qualitative comparisons among those papers. These works are summarised in Table 2 and are never discussed in detail in the text. The paper focused on the comparison of FL architectures instead which has been widely considered by many previous surveys as stated by the paper.
We put three more papers [82] and [97-98] to the use of FL in AQI. We discussed in text at line 72-167
4- In table 2, what are evaluation meters? is it metrics instead?
Yes, we replaced meters with metrics in table 2. Please see it at the top of 4th page.
5- In Table 2, the works presented, How the energy consumption reduced? These works are questionable in terms of their claim as their energy consumption numbers do not seem realistic and are not based on accurate energy models or real measurements. So it is not clear whether the conclusions made are appropriate.
We remove the energy consumption in table 2.
6- The heterogeneity challenge in line 231, Lacks citations of works that studied the heterogeneity impact such as https://dl.acm.org/doi/abs/10.1145/3517207.3526969 or https://dl.acm.org/doi/abs/10.1145/3442381.3449851 and another work that leveraged adaptive model quantization to address it such as https://dl.acm.org/doi/abs/10.1145/3437984.3458839
Yes, we cited the works that you mentioned. They are in [94-96], please see them at line 294. Thank you so much
7- In future works, Not clear what is the motivation for multi-model esp. for the AQI use-case. The authors cite works that perform Multimodal learning in AQI prediction use-case but this has no relevance to the proposed multi-model FL approaches. It seems the authors also fail to make the distinction clear between multi-modal, multi-model, and multi-task learning.
Yes, we use another paper for the citation [74] “Bhuyan, N., Moharir, S., & Joshi, G. (2022). Multi-Model Federated Learning with Provable Guarantees. ArXiv, abs/2207.04330”, please see it at line 494. We also replaced the citation [42-43] to the multi-model instead of multimodal. They are now at line 443. Thank you
8- The work concludes that multi-model federated learning is the best approach without giving any solid observations or supporting evidence. It would be beneficial if some experimental evaluations are used to support the claims.
We rewrote the abstract that concludes multi-model federated learning is the best approach, please see the next response. Besides that, we replaced the words “the best fit” with “applied”, it is now at line 509. Thank you
9- Finally, the authors make an over-claim by stating that they surveyed 90 papers in the scope of AQI predictions which is not the case. Having 90 citations in the work does not mean you have surveyed them all.
=> We rewrote the abstract: “We have examined over 90 carefully selected papers in this scope and discovered that multi-model federated learning is the most effective technique that could enhance the air quality index prediction result. Therefore, this mechanism needs to be considered by science community in the coming years” (line 7 - 10) TO “In this survey, we went over the works that previous scholars have done in AQI forecast both in traditional machine learning (ML) approaches and federated learning (FL) mechanisms. Our objective is to comprehend previous research on air pollution prediction including methods, models, data sources, achievements, challenges, and solutions applied in the past. We also convey a new path of using multi-model federated learning, which has piqued the computer science community’s interest recently”
Author Response File: Author Response.pdf
Round 2
Reviewer 1 Report
The article is in better shape. Most of the changes have been incorporated into the current version. However, the referencing is still not in order. E.g. last line on the Page is reference 7 (the last one). But on page two the first reference in table 1 starts from 45. It should be started from 8 and onwards continued.
Right now reference 8 is on page 3. It should be continuous numbering in order from 1,2,3..........
Author Response
REVIEWER 1 ROUND 2
Dear Reviewer. Thank you for your comments. We appreciated them. We adjusted the paper as follows:
The article is in better shape. Most of the changes have been incorporated into the current version. However, the referencing is still not in order. E.g. last line on the Page is reference 7 (the last one). But on page two the first reference in table 1 starts from 45. It should be started from 8 and onwards continued.
Right now reference 8 is on page 3. It should be continuous numbering in order from 1,2,3..........
Yes. We reordered all the references. Thank you so much.
Author Response File: 
Author Response.docx
