DESTformer: A Transformer Based on Explicit Seasonal–Trend Decomposition for Long-Term Series Forecasting
Round 1
Reviewer 1 Report
Thank you for the opportunity to review.
This is a useful and original contribution to the field.
The following chapters clearly describe the course of research, suggest a new method that will be very interesting to learn the reader. The manuscript itself is well structured, figures and tables are presented and described in full.
Here I provide my remarks.
Authors must use a template.
All section and subsection must be numbered.
It would be better to added at the abstract: highlight the purpose of the study, summarize the articles main findings and indicate the main conclusions or interpretations.
Abstract and conclusion must be supported by data.
All Figures and Tables musts be cited and described in the text before you show figure or table. See all text.
Figure 3 is of poor quality. Text unreadable.
In my opinion, the authors should expand the conclusion section. You provide extensive and interesting research and have concluded it in five sentences, without even supporting the data, ignoring the lack of discussion.
Improve References. Add more the cited work from Applied Sciences journal.
The manuscript could be accepted after improve.
Author Response
Please see the attachment.
Author Response File: 
Author Response.docx
Reviewer 2 Report
In my humble opinion, the article reads like an outline for a 15-minute presentation at a conference with people in the field. Because several pre-algorithms are simply cited, several symbols are not defined in the text and several disadvantages of previous methods have been pointed out, it is not clear how the new proposal solves these problems. Furthermore, it must be said that only empirical evidence is presented that the new method is better than the previous ones, as an efficiency comparison is shown on some data series. The article also does not present any proof that guarantees the convergence as well as the order of magnitude of the complexity of the new method, is it O(Llog(L))??
More specifically:
1) For me it is very important to see the algorithm of the new method: DSTformer . With figure 1 alone it is not clear how DSTfortmer works, that is, if I want to test this algorithm only with the provided architecture, it is not possible to write the code.
2)How DSTformer handles: " as sensitivity to outliers and missing values, difficulty in dealing with nonlinear and complex time series data, and difficulty in integrating other relevant information" ??
3)What is the complexity of DSTformer, O(LlogL)? can you prove it?
4) In line 115 : "In this paper, we introduced the idea of decompo- 115 sition from a new progressive dimension" What is " new progressive dimension" ??
5) What is : P(Y|HS, HT) ? Is it the probability of predicting a new series given seasonal and trend components?
6) Lines 149,150 : "a new sequence decomposition method is used that maps the sequence to the frequency domain and takes high frequencies as seasonal components and low frequencies as trend components" Who decides what is high or low frequency? Is it a subjective choice? And how to treat white noise?
7)Please explain equation 1 a lot more, it is an enigma!! What is M and why M=3?
8)Please define MSE and MAE!! I would like to see new time series with DSTformer predictions instead of two simple statistics measures that can dilute large errors!!
For me the text is well written in the English language.
Author Response
Please see the attachment.
Author Response File: Author Response.docx
Reviewer 3 Report
In this manuscript, authors proposed a new transformer (DSTformer) which can decompose long-term series forecasting into seasonal and trend components. Even though the manuscript provides enough experimental analysis and results, some questions need to be solved as follows:
1. Naming problem: DSTformer. The title of “DSTformer” already used by other authors’ article may be changed into the title of the revised manuscript. Authors can check out the problem as following: https://arxiv.org/pdf/2210.06551v1.pdf
2. Authors should reorganize the total structure of the manuscript using Sections and sub-Sections.
3. In Related Work, authors focus on time complexities (O(LlogL)) of the previous studies. However, authors insist on the low time complexity such as O(L). Please provide the proof of it as plot or others.
4. Please explain that why authors select Autoformer instead of Fedformer for comparison, in Figure 3. If possible, please add the T-SNE of Fedformer in Figure 3.
5. Please update typos, abbreviations, and italic expressions in the revised manuscript.
The level of English language shown in the manuscipt is not bad.
Authors need to update the manuscript in alignment with MDPI format totally.
Author Response
Please see the attachment.
Author Response File: Author Response.docx
Round 2
Reviewer 1 Report
Thank you for the comments. Tha manuscript can be accepted. Unfortunately, the authors did not improve the quality of Figure 3, the text in the figure is unreadable (ok, blue and red you designate in the caption of figure, but the numbers on the axis of coordinates are unreadable).
Author Response
addressed
Reviewer 2 Report
The authors answered most of the questions raised in the first review and improved the text. Therefore, I am in favor of publication.
Author Response
addressed
