1) With all due respect to authors that write technical text in a non-native language, there are nevertheless certain minimal standards that have to be met. This work fails in this respect, and therefore the English style and grammar should be improved as much as possible. I would recommend asking a colleague for help. 2) It would also improve the paper if the table and figure captions would be made more self contained. In addition to what is shown, one could also consider a sentence or two saying what is the main message of each figure and data in the table. 3) The presentation is not in keeping with an interdisciplinary readership. Too many technical details are presented without much guidance to the reader through what is shown and why. This needs major improvement for better clarity of the presentation. 4) For a better introduction of the background and related work, the authors could cite also recently published related papers and reviews, namely: Wavelet entropy-based evaluation of intrinsic predictability of time series, Ravi Kumar Guntu, et al., Chaos 30, 033117 (2020) and Signal propagation in complex networks, Peng Ji, et al., Phys. Rep. 1017, 1-96 (2023). This is related and relevant for the subject of the present paper. 5) Some references contain errors, missing or incorrect information, and inconsistent formatting. It is difficult to give credit to research if such elementary aspects of the work are not error free. References should thus be corrected with the best care. In particular, the authors should use the article number for journals where page numbers are no longer used. 6) Finally, I would encourage the authors to shorten the abstract and focus more on key results. As it is, the abstract is verbose and with a lot of trivial and partly also disputed information, and the main messages dilute in such a style. # Original Copy 03-Jan-2024 Dear Mr. TSUBOUCHI: I am writing to you regarding manuscript # Access-2023-41159 entitled "MetricSifter: Feature Reduction of Multivariate Time Series Data for Efficient Fault Localization in Cloud Applications" which you submitted to IEEE Access. Your article was peer reviewed with interest but has not been recommended for publication in its current form. We strongly encourage you to address the reviewers’ concerns, which can be found at the bottom of this letter, and resubmit your article to IEEE Access once you have updated it accordingly. Please note that IEEE Access has a binary peer review process. Therefore, to uphold quality to IEEE standards, an article is rejected even if it requires minor edits. When updating your manuscript, you should elaborate on your points and clarify with references, examples, data, etc. If you disagree with any technical points the reviewers have made, please include your counterarguments in your response to the reviewers (more information detailed below) and work this into the updated manuscript. Also, note that if a reviewer suggested references, you should only add those that are relevant to your work if you feel they strengthen your article. Recommending references to specific publications is not appropriate for reviewers and you should report excessive cases to [[email protected]](mailto:[email protected]). If the updated manuscript is determined not to have addressed all of the previous reviewers’ concerns, or if the Associate Editor still has substantial technical concerns, the article may be rejected and no further resubmissions will be allowed. When you are ready to resubmit your updated article, you can do so in the IEEE Author Portal. When you log into the IEEE Author Portal you ill see the title of the rejected article and the option to “Start Resubmission”. Upon resubmission you will be asked to upload the following 3 files: 1. A document containing your response to reviewers from the previous peer review. The “response to reviewers” document (template attached) should have the following regarding each comment: a) Reviewer’s concern, b) your response to the concern, c) your action to remedy the concern. The document should be uploaded with your manuscript files under "Author's Response Files.” 2. Your updated manuscript with all your individual changes highlighted, including grammatical changes (e.g. preferably with the yellow highlight tool within the pdf file). This file should be uploaded with your manuscript files as “Highlighted PDF.” 3. A clean copy of the final manuscript (without highlighted changes) submitted as a Word or LaTeX file, and as a PDF, both submitted as the “Main Manuscript.” **IMPORTANT: Please see the attached Resubmission Checklist that details all the items listed above. Please utilize this checklist to ensure you have made the necessary edits to your manuscript, and to ensure you have all the necessary files prepared prior to resubmission. *** AUTHOR LIST CHANGES: If your revised manuscript has an updated author list, you will need to submit a formal request to the Editor by completing the attachment labelled ‘Request for Byline Change,’ and uploading it as 'Request for byline change form.' This should include a DETAILED justification explaining each author’s contribution(s) to the work. Change in the author list is considered rare and exceptional, and the decision to allow such changes rests with the Editor. Once the list and order of authors has been established, the list and order of authors should not be altered without permission of all living authors of that article. We sincerely hope you will update your manuscript and resubmit soon. Please contact me if you have any questions. Thank you for your interest in IEEE Access. Sincerely, Prof. Yang Liu Associate Editor, IEEE Access [[email protected]](mailto:[email protected]), [[email protected]](mailto:[email protected]) ## Reviewers' Comments to Author: ### Reviewer: 1 Recommendation: Accept (minor edits) Comments: In this paper, the authors introduce MetricSifter, a novel framework designed to improve fault localization by efficiently reducing irrelevant features. Utilizing a binary classification approach, MetricSifter distinguishes between failure-related and unrelated monitoring metrics, with a focus on the temporal proximity of fault-induced changes. The framework combines offline change point detection and probability density estimation to pinpoint the most significant change points. Results from simulations and empirical tests demonstrate MetricSifter's superior performance over existing methods. Additional Questions: Please confirm that you have reviewed all relevant files, including supplementary files and any author response files, which can be found in the "View Author's Response" link above (author responses will only appear for resubmissions): Yes, all files have been reviewed 1. Does the paper contribute to the body of knowledge?: Yes 2. Is the paper technically sound?: Yes 3. Is the subject matter presented in a comprehensive manner?: Yes 4. Are the references provided applicable and sufficient?: Yes 5. Are there references that are not appropriate for the topic being discussed?: No 5a) If yes, then please indicate which references should be removed.: ### Reviewer: 2 Recommendation: Reject (updates required before resubmission) Comments: In this study, the challenges associated with fault localization in large-scale cloud-based applications are addressed due to the intricate interdependencies among their components and the abundance of operational data. Recent research endeavors have concentrated on automating fault localization by leveraging statistics and machine learning to mine time series data of monitoring metrics. To enhance the accuracy of localization, fault localization methods incorporate feature reduction techniques aimed at reducing the number of monitoring metrics unrelated to failures. However, these methods grapple with inaccuracies stemming from either excessive reduction of failure-related metrics or retention of too few failure-unrelated metrics. This paper introduces MetricSifter, a feature reduction framework explicitly designed to identify anomalous metrics resulting from faults with precision. The key insight driving this framework is the observation that change points within monitoring metrics tend to cluster in time during the failure duration. Based on this insight, the algorithm is devised to pinpoint a time segment exhibiting the highest density of change point occurrences across monitoring metrics. Simulation results demonstrate that MetricSifter achieves a reduction accuracy of 0.981, surpassing the best baseline method by an accuracy margin of 0.178. Moreover, experiments conducted using datasets generated from two open-source benchmark applications provide further evidence that feature reduction can enhance fault localization performance. Importantly, the results obtained through MetricSifter outperform those of the baseline methods. But while I appreciate the subject being studied, I also have comments that require revision with care and love to detail. 1) With all due respect to authors that write technical text in a non-native language, there are nevertheless certain minimal standards that have to be met. This work fails in this respect, and therefore the English style and grammar should be improved as much as possible. I would recommend asking a colleague for help. 2) It would also improve the paper if the table and figure captions would be made more self contained. In addition to what is shown, one could also consider a sentence or two saying what is the main message of each figure and data in the table. 3) The presentation is not in keeping with an interdisciplinary readership. Too many technical details are presented without much guidance to the reader through what is shown and why. This needs major improvement for better clarity of the presentation. 4) For a better introduction of the background and related work, the authors could cite also recently published related papers and reviews, namely: Wavelet entropy-based evaluation of intrinsic predictability of time series, Ravi Kumar Guntu, et al., Chaos 30, 033117 (2020) and Signal propagation in complex networks, Peng Ji, et al., Phys. Rep. 1017, 1-96 (2023). This is related and relevant for the subject of the present paper. 5) Some references contain errors, missing or incorrect information, and inconsistent formatting. It is difficult to give credit to research if such elementary aspects of the work are not error free. References should thus be corrected with the best care. In particular, the authors should use the article number for journals where page numbers are no longer used. 6) Finally, I would encourage the authors to shorten the abstract and focus more on key results. As it is, the abstract is verbose and with a lot of trivial and partly also disputed information, and the main messages dilute in such a style. If a revision is granted, I will be happy to review the manuscript again. Additional Questions: Please confirm that you have reviewed all relevant files, including supplementary files and any author response files, which can be found in the "View Author's Response" link above (author responses will only appear for resubmissions): Yes, all files have been reviewed 1. Does the paper contribute to the body of knowledge?: yes 2. Is the paper technically sound?: partly 3. Is the subject matter presented in a comprehensive manner?: partly 4. Are the references provided applicable and sufficient?: partly 5. Are there references that are not appropriate for the topic being discussed?: No 5a) If yes, then please indicate which references should be removed.: If you have any questions, please contact article administrator: Ms. Namrata Sinha [[email protected]](mailto:[email protected])