Bi-PredRNN: An Enhanced PredRNN++ with a Bidirectional Network for Spatiotemporal Sequence Prediction
Round 1
Reviewer 1 Report
Comments and Suggestions for AuthorsThe manuscript is well-organized and demonstrates a clear structure and logical flow. The research appears technically sound, but several aspects require clarification and improvement to ensure the study’s relevance and accessibility to the target audience of the journal.
1) While the paper focuses on improving a specific algorithm, it is unclear whether this focus aligns well with the scope of the journal.
2) The expression style in the manuscript is often complex and difficult to follow, even for readers experienced in machine learning research. Simplifying the technical descriptions and improving the clarity of expressions would significantly enhance the readability of the paper.
3) The manuscript evaluates the proposed model using four examples, but the presentation of the data is insufficiently detailed. Incorporating visual figures to display example data would improve the accessibility of the cases.
4) Additionally, the authors should clarify the number of data points used for training, validation, and testing in each case, as the current explanation is vague and leaves readers uncertain about the data partitioning process. There are inconsistencies in the dataset splitting strategies across examples. For instance, the Moving MNIST dataset is split into 10,000 training sequences and 3,000 validation sequences, whereas for the KTH dataset, the split is organized into 16 training sequences, 4 validation sequences, and 5 test sequences. The rationale for these differing splits is not explained and should be addressed.
5) In the results section, the focus is primarily on overall metrics for model evaluation. Adding visualizations of the prediction process on test sets would provide valuable insights. This is particularly important for identifying model performance at critical points, such as sudden changes, which are often of greater practical significance.
Comments on the Quality of English LanguageThe expression style in the manuscript is often complex and difficult to follow, even for readers experienced in machine learning research. Simplifying the technical descriptions and improving the clarity of expressions would significantly enhance the readability of the paper.
Author Response
Dear Editors and Reviewer, Thank you for your thorough and constructive review of our manuscript. We deeply appreciate the time and expertise you have dedicated to providing valuable feedback that has significantly improved our paper. We have addressed all comments and suggestions in the attached point-by-point responses. Please refer to the attachment for our detailed responses to each reviewer's comments and the corresponding revisions made to the manuscript. We believe these revisions have substantially strengthened the paper, and we look forward to your further evaluation.
Author Response File: 
Author Response.pdf
Reviewer 2 Report
Comments and Suggestions for AuthorsIt is a very interesting paper entitled Step Increment Date, undoubtedly a valuable contribution serving several purposes for the literature. Nonetheless, there are some areas to improve toward enhancing its quality and impact in general. Equations (2-4): It would be better to revise Equations (2-4) since it is not explicit whether these summations are indeed supposed to be equal to that second part. A clear explanation or proof regarding this equivalence would make the mathematical framework more robust. Table 1 - Key Results: The results presented in Table 1 are what this paper is all about; however, without a statistical test to support it, the results will lose some of their intrinsic power. I recommend the authors include some kind of statistical test. One of which can be the Kolmogorov-Smirnov Based Test for Comparing the Predictive Accuracy of Two Sets of Forecasts. This can be done using the available R package and would prove whether or not the observed improvements are statistically significant. An addition of the graph that demonstrates the sensitivity of the best-performing model to noise levels would make a big difference to the paper. A well-drawn graph could give the reader an easy view of how well the model does in different conditions and therefore add some real depth to the analysis. Simulation for Model Robustness: Further, a cross-sectional simulation would add value from a variety of datasets. It would also achieve an important aspect of the comparative performance of the model’s sensitivity in those areas where other models do not perform well, but the proposed model does. This would provide very strong evidence of the model’s practicability and generalization ability.
Author Response
Dear Editors and Reviewer, Thank you for your thorough and constructive review of our manuscript. We deeply appreciate the time and expertise you have dedicated to providing valuable feedback that has significantly improved our paper. We have addressed all comments and suggestions in the attached point-by-point responses. Please refer to the attachment for our detailed responses to each reviewer's comments and the corresponding revisions made to the manuscript. We believe these revisions have substantially strengthened the paper, and we look forward to your further evaluation.
Author Response File: Author Response.pdf
Reviewer 3 Report
Comments and Suggestions for AuthorsThere are several aspects that need to be improved to make the manuscript clearer and more comprehensive. The main points are outlined below:
1. Provide a more precise explanation of how the forecasting horizon k is integrated into the model and what possible boundary conditions are considered.
2. Include more details about the "dynamic window size adaptation" mechanism, as this term is mentioned superficially in the introduction.
3. While Bi-PredRNN is motivated by bi-directionality, the introduction lacks a more detailed explanation of why this advantage is significant compared to other methodologies. A theoretical comparison with unidirectional models is recommended.
4. The literature review is comprehensive, but some references (e.g., [1–4], [16]) appear to be formally cited without a thorough explanation of their application context. It is recommended to provide a brief description of these sources' contributions to the development of Bi-PredRNN.
5. Considering that Section 2 contains a lot of technical details, it is suggested to remove secondary information (e.g., the description of DropPath) and focus on the technologies or methods directly related to the proposed model.
6. It is recommended to include a clearer transition between the literature review and the proposed method. For example, highlighting how the Bi-PredRNN architecture addresses specific limitations mentioned in the analysis of current models would be beneficial.
7. While the mathematical models presented are detailed, some formulas could be supplemented with explanations (e.g., why certain loss coefficients were chosen or how the sigmoid function is specifically applied in window size calculation).
8. Why was CausalLSTM specifically chosen? Was its performance compared with other alternatives (e.g., GRU, Transformer)?
9. The description of the dynamic window mechanism is technically accurate but somewhat complex for the reader. Concrete examples of how different motion intensities influence window size changes could be provided.
10. While the "Adam" optimizer and a fixed learning rate of E−3 were mentioned, information on other important hyperparameters, such as the number of layers, latent dimension size, or dropout rates, is missing. Suggestion: Describe the key hyperparameters and the optimization process (e.g., was hyperparameter search performed, and if so, by what method?).
11. Data preprocessing to remove seasonal cycles and other systematic variations (especially for climate data) was mentioned, but there is no discussion on whether such preprocessing might have influenced model training. Suggestion: Add insights into whether this preprocessing could have affected the model’s ability to generate long-term predictions.
12. The metric definitions provided are clear, but there is no discussion of why these specific metrics were chosen or how they are interpreted in analyzing the model’s strengths and weaknesses.
13. In the results section, the idea that the bi-directional architecture and dynamic window size mechanism are the key elements driving improvement is repeated multiple times. This could be presented more concisely.
14. It would be useful to present alternative validation strategies (e.g., cross-validation or randomized test sets) and discuss how these could reduce the risk of overfitting. Additionally, it is worth suggesting an extension of the analysis with regularization methods to minimize the impact of correlations.
15. The provided metrics (MSE, PSNR, SSIM) are not supported by variance analysis or statistical significance testing, making it unclear whether the differences between models are statistically significant.
16. While the conclusions highlight the model's efficiency, a broader discussion of real-world applications and possible use cases where this method could be beneficial is missing.
17. The problem that future states influencing the past can introduce uncertainty is mentioned, but a more detailed explanation of how and why this occurs is lacking.
18. The proposed improvements are interesting but lack detailed information on their integration into the existing architecture. The conclusions could be supplemented with a brief description of how the proposed mechanisms (e.g., temporal attention) could be technically applied to the bi-directional memory architecture.
Author Response
Dear Editors and Reviewer, Thank you for your thorough and constructive review of our manuscript. We deeply appreciate the time and expertise you have dedicated to providing valuable feedback that has significantly improved our paper. We have addressed all comments and suggestions in the attached point-by-point responses. Please refer to the attachment for our detailed responses to each reviewer's comments and the corresponding revisions made to the manuscript. We believe these revisions have substantially strengthened the paper, and we look forward to your further evaluation.
Author Response File: Author Response.pdf
Round 2
Reviewer 1 Report
Comments and Suggestions for AuthorsThe manuscript has undergone comprehensive revisions and has been significantly improved. All concerns have been addressed clearly, making it easier for readers to follow and understand. I recommend the manuscript for acceptance.
Reviewer 2 Report
Comments and Suggestions for AuthorsI have no further comments.
Reviewer 3 Report
Comments and Suggestions for AuthorsThe authors' response demonstrates attention to the provided comments and efforts to improve the manuscript.
Overall, the authors have shown an understanding of the provided comments and have successfully integrated them into the manuscript. In my opinion, the manuscript has been considerably improved compared to the initial version and is close to meeting publication requirements.
