Review Reports - Pm2.5 Time Series Imputation with Deep Learning and Interpolation

Round 1

Reviewer 1 Report

The paper addresses the issue concerning the NA values in time series.

The issue addressed by the authors is essential since NA values are a big issue in the time series analysis. However, the paper needs more details and description. The novelty description is naïve, and it is challenging the understand the research behind the paper. The authors have made an interesting effort in terms of implementation and experiments, but they lack words describing the novelty they propose.

The related work should be improved, including more examples of time-series analysis and evaluation such as (DOI: 10.1109/ICASSP43922.2022.9746342, DOI: 10.4108/icst.bodynets.2013.253551, DOI: 10.1109/LSP.2022.3224880 , DOI: 10.1109/IE.2013.12). A final paragraph summary of the differences between the proposed paper and the related world would be appreciated.
The proposed approach must be described better. The statement: "The imputation problem is approached as a classification problem, as well the proposal uses a classification model to determine the class of NA value and, from the class, an interpolation technique is implemented." It must be highlighted in the paper. A flowchart could be helpful.
Improve the quality of Figure 2, and Figure 3 and provide a more descriptive title.
"3.1. Data Preparation" more details, please. Have the data de-noised? Normalized? Please provide an example of data.
About Table 1, how do the authors get to these features? What is the motivation for this selection?
There are two "Figure 3" please fix it.
"3.3.3 Interpolation according to class estimation" This paragraph must be detailed better since it is supposed to present the paper's novelty.
"4.1 Results" The authors should clearly describe the results to prove the goodness of their work.

English should be improved.

Author Response

Dear Reviewer, manuscript has been revised according to provided recommendations.

Recommendation	Action
The related work should be improved, including more examples of time-series analysis and evaluation such as (DOI: 10.1109/ICASSP43922.2022.9746342, DOI: 10.4108/icst.bodynets.2013.253551, DOI: 10.1109/LSP.2022.3224880, DOI: 10.1109/IE.2013.12). A final paragraph summary of the differences between the proposed paper and the related world would be appreciated.	Related works 10.1109/ICASSP43922.2022.9746342 and 10.1109/LSP.2022.3224880 have been added to the manuscript. However, 10.1109/ICASSP43922.2022.9746342 and 10.4108/icst.bodynets.2013.253551 are not time series imputation-related works so they were discarded. A final paragraph and Figure 2 have been added, both contain the differences between the proposal approach and literature approaches. (See pages 3 and 4)
The proposed approach must be described better. The statement: "The imputation problem is approached as a classification problem, as well the proposal uses a classification model to determine the class of NA value and, from the class, an interpolation technique is implemented." It must be highlighted in the paper. A flowchart could be helpful.	To highlight, it was emboldened (See page 1). Also, Figure 2 (added, see page 4) and Figure 3 (improved, see page 4), both present a graphical view of the proposal.
Improve the quality of Figure 2, and Figure 3 and provide a more descriptive title.	Figure 2 now Figure 3 (See page 4), and Figure 3 now Figure 5 (See page 7), both have been improved and have a more descriptive title.
"3.1. Data Preparation" more details, please. Have the data de-noised? Normalized? Please provide an example of data.	The data was normalized, this process is described in section 3.2.3. (See page 8) An example of data is provided in Figure 4 (See page 5)
About Table 1, how do the authors get to these features? What is the motivation for this selection?	The motivation was added (See page 6)
There are two "Figure 3" please fix it.	Issue was fixed
3.3.3 Interpolation according to class estimation" This paragraph must be detailed better since it is supposed to present the paper's novelty.	This section was improved. The algorithm for the polynomial procedure in Figure 12 was added. Also, the algorithm interpolate procedure in Figure 13 was added. (See page 15) including respective descriptions.
"4.1 Results" The authors should clearly describe the results to prove the goodness of their work.	Results section has been improved (See page 19)

Reviewer 2 Report

The paper presents the nice idea with a clear way.

I have only two comments:

1. Does this code outperforms simple techniques that address the NA problem? (e.g. the average of previous and next obs)

2. Can you add an appendix providing the whole code?

Author Response

Dear Reviewer:

The revised manuscript contains the improvements according to provided recommendations. These are highlighted in yellow color.

Recommendation

Action

1. Does this code outperforms simple techniques that address the NA problem? (e.g. the average of previous and next obs)

The technique that averages previous and next observations is known as Local Average of Nearest Neighbors (LANN) and it is included in this work as a benchmark model. (See pages 2, 17,18,19)

2. Can you add an appendix providing the whole code?

As it is a long code, a link to download was provided

https://drive.google.com/drive/folders/1qL-k80rXqjFVmi-4fubg6nFvQbj3-Xq6

(See page 16)

Reviewer 3 Report

The authors present a topic of interest, namely the processing of time series using Deep Leraning with interpolation. The authors demonstrate the concept by proposing several methods of both neural networks and statistical algorithms, the results being conclusive.

I ask the authors to make a comparison of the algorithm proposed in the article with similar articles from the specialized literature, but on the subject of time series analysis with deep learning and fractal theory.

Fractal theory is expanding compared to traditional algorithms (eg Machine Learning in Classification Time Series with Fractal Properties).

Author Response

Dear Reviewer

The provided recommendation has been taken in count as future work (See page 22) due that Fractal Theory has not been used yet for time series imputation. It could help to find another kind of features that could improve the metrics of classification models used in the proposal.