Time Series Forecasting Performance of the Novel Deep Learning Algorithms on Stack Overflow Website Data
Round 1
Reviewer 1 Report
The authors need to address the following issues:
1. Only the number and types of layers in each model are described without specifying the number of neurons or units in each layer is should be specified in table 1.
2. Accuracy of time-series prediction usually decreases as the forecast horizon increases. The authors may also evaluate the models on this metric. as well for better comparison.
Author Response
Thank you for the comments and suggestions.
- We added a new table to provide more detail about models (layers, units, execution time per epoch, etc.),
- You mentioned an important general tendency for time series prediction that prediction accuracy decay if the prediction window is increased. So, to display this property, we added new results for 5-day-long predictions.
Author Response File: Author Response.docx
Reviewer 2 Report
In this paper, the performance of a new deep learning algorithm for predicting the time series of stack overflow website data is studied,The topic is interesting.This paper present that wavenet-based CNN models can also extract useful knowledge from datasets
with the dimension of time and yield more precise prediction results than LSTM models. It is necessary to supplement the universal method performance test of the proposed algorithm and conduct preliminary theoretical analysis
Author Response
Thank you for the comments and suggestions.
You suggested supplying universal test results for the proposed algorithm. We used the S&P 500 dataset for evaluating the prediction performance of the wavenet. It did perform very well while other models failed. And, for future works, wavenet model performance can be compared by using the same dataset (S &p 500)
Author Response File: Author Response.docx
Reviewer 3 Report
This research focuses on deep learning models for evaluating Time Series Forecasting Performance and Dense Model, Stateless RNN Model, Stateful RNN Model, LSTM Model, CNN Hybrid Model and Fully CNN Model are selected.
Consider applied research, the current presentation is in good condition. Some recommendations here can improve on this aspect,
1. Authors highly suggest adding some Literature Review (their algorithm) on those selected deep learning models. This will help novios readers for better understanding.
2. Authors also suggested providing a general flow chart for representing flows/steps on these selected deep learning models.
With the above points, I suggest ACCEPTING this paper by considering the stated recommendations.
Author Response
An extra part is added to the end of the literature for providing some preliminary information and literature about the used algorithms.
Author Response File: Author Response.docx
Reviewer 4 Report
The research presents valuable insights on univariate time series main deep learning algorithms and introduces a clear and structure definition of mathematical and technical background of the algorithms. However, some modifications are suggested to improve technical quality of the paper.
In section 4 the description of used dataset is given, more details about preprocessing phases are needed,
In results discussion section it will be relevant to include execution time as well for the compared models, for real use cases in industry this can be a relevant metric.
The paper deals with univariate time series forecasting , based on conclusions from this work it would relevant to give insights about multivariate times series forecasting through deep learning that has some less detailed explanation and use in the literature
Author Response
Thank you for the suggestions; However, some modifications are suggested to improve the technical quality of the paper.
- more details about preprocessing phases are added; "The data is extracted from the website by the corresponding API and saved as a .csv file. This file is read and processed by the pandas’ data frame method. And hourly frequency of the asked questions is used as the time series data. The property of this data is also added since it is an IID set and created in a real-world example."
- it will be relevant to include execution time; In Table 3, the execution time and complexity of each model are added.
- give insights about multivariate time series forecasting through deep learning; In the related work part, an extra paragraph is added to address the multivariate time series prediction and provide some literature.
Author Response File: Author Response.docx
Round 2
Reviewer 1 Report
The revised manuscript may be accepted for publication.