投稿TMC的感受

投稿任务告与段落了，最终的结果是被TMC给early reject了。这神一样的审稿意见让我真的是老头地铁看手机啊！所以虽然TMC没有给我rebuttal的机会。所以我还是打算在CSDN进行一次rebuttal。
其实我做的东西很简单，就是把时间序列中的将时间序列转化为图的思想应用到无线定位领域。以下是TMC的审稿意见：
Editor Comments

Associate Editor
Comments to the Author:
For your paper I got three independent reviews from experts in various fields. All the Reviewers made a plethora of major and minor concerns. After my own reading, my concerns are as follows:

• Relevant papers from the state of the art are not cited, discussed, and comparisons with the submitted results is missing.（这开局既是王炸啊！那按照副主编的意思我后面的对照实验和消融实验都是白做的，另外论文里面50几篇参考文献都是白引用的呗。）
• The novelty is questionable since concepts such as “inception structure” and “residual connection” were introduced many years ago.（无敌，这个审稿意见，笑死我了，这是我写的创新点，The inception structure and residual connection are used to horizontally expand the correlation feature
  aggregation structure, thereby avoiding the over-smoothing issue of traditional graph neural network。这是副主编的审稿意见:2) The inception structure and residual connection approach is not novel, contrary to what is stated in section 1.4. The inception structure was first introduced in the context of convolutional neural networks (CNNs) by the Google research team in 2014, while the residual connection was proposed in 2016 as a way to address the vanishing gradient problem in deep neural networks.真的这意见甚至不如我拿chatpaper生成的来的客观准确啊！真的这审稿意见简直无敌了，一个图神经网络的论文，三个意见包括副主编的意见居然和图神经网络没有任何关系。话说你一个副主编把审稿意见写成这样你的良心真的不会痛吗？而且作为一个在学术圈有头有脸的人物写出这样的意见，你真的不该感受到害臊吗？下面是我用chatpaper生成的本文主要创新点：
  ）
• The methodology of the paper is not clear as correctly noted by the 1st Reviewer.（这个我们等一下看审稿人1的奇葩意见）
• The adoption of LSTM architecture is not justified, and the training overhead is not discussed.（啊！ justified牛啊，直接用了这个单词，太牛逼了，但是我想问问LSTM怎么就不合适了，那么多动态图的算法不都是LSTM+GNN吗？比如Graph LSTM、GC-LSTM等等等等。怎么到你这就不合适了？至于讨论训练开销？我估计在TMC副主编的狗眼里面，消融实验不算讨论。）
  Considering all the above, the paper lacks sufficient novelty and technical depth to be published in IEEE TMC.（综上跟条狗一样）

---

Reviewer Comments

Please note that some reviewers may have included additional comments in a separate file. If a review contains the note “see the attached file” under Section III A – Public Comments, you will need to log on to ScholarOne Manuscripts to view the file. After logging in, select the Author Center, click on the “Manuscripts with Decisions” queue and then clicking on the “view decision letter” link for this manuscript. You must scroll down to the very bottom of the letter to see the file(s), if any. This will open the file that the reviewer(s) or the associate editor included for you along with their review.

Reviewer: 1

Recommendation: Author Should Prepare A Major Revision For A Second Review

Comments:

1. The proposition of this work is fairly new, and the paper is well-written in general.
2. The inception structure and residual connection approach is not novel, contrary to what is stated in section 1.4. The inception structure was first introduced in the context of convolutional neural networks (CNNs) by the Google research team in 2014, while the residual connection was proposed in 2016 as a way to address the vanishing gradient problem in deep neural networks（绝，绝种绝，看来和副主编的意见一模一样啊！这是我写的创新点，The inception structure and residual connection are used to horizontally expand the correlation feature
   aggregation structure, thereby avoiding the over-smoothing issue of traditional graph neural network。这是副主编的审稿意见:2) The inception structure and residual connection approach is not novel, contrary to what is stated in section 1.4. The inception structure was first introduced in the context of convolutional neural networks (CNNs) by the Google research team in 2014, while the residual connection was proposed in 2016 as a way to address the vanishing gradient problem in deep neural networks.真的这意见甚至不如我拿chatpaper生成的来的客观准确啊！真的这审稿意见简直无敌了，一个图神经网络的论文，三个意见包括副主编的意见居然和图神经网络没有任何关系。话说你一个副主编把审稿意见写成这样你的良心真的不会痛吗？而且作为一个在学术圈有头有脸的人物写出这样的意见，你真的不该感受到害臊吗？下面是我用chatpaper生成的本文主要创新点：这个审稿意见类似于我说前门楼子，你说胯骨轴子啊！而且我很奇怪审稿人为什么不给我找一篇用残差和inception解决过平滑的文章来打我脸呢？而是找了inception和残差的原始论文呢？别告诉我审稿人不知道什么是过平滑啊！）.
3. It is not clear why the authors want to use time-frequency analysis on the signal graphs? Couldn’t the correlations be taken directly from the identified features since ML techniques are being used? The need and the outcome should be clarified before the detailed explanations in section 2.2 and 2.34) What type of verification is done using Dataset 1 as opposed to Datasets 2 & 3?（我真的词穷了，那GCN不就是图傅里叶变换吗？那图拉普拉斯矩阵的特征值不就是图的频率吗？来你告诉我这不叫时频这叫什么，这叫什么？？无语！！！至于“既然使用了ML技术，就不能直接从已识别的特征中提取相关性吗？”当然能啊！就是效果不好，比较实验你不会没看吧！至于为什么用多个数据集验证，那验证模型的鲁棒性不就是用多个数据集验证吗？咋了？到你这儿不行了是吗？）
4. A discussion must be included on the results. Why the proposed algorithm performs better in an unsupervised regime as compared to supervised, in case of varying environments?（哎呀我的妈！有监督和无监督在本文里面是不同的case study，只是说本文的算法可以适用于无监督的情况。至于说什么无监督比有监督的好，那条件设置和评价标准都不一样你是怎么看出来无监督比有监督的好呢？不会就看了看数吧？实验设置和评价标准完全没看？）
5. It is advised that the authors reduce the length of theoretical part in first 4 pages of the manuscript and explain better the methdology and results. Otherwise, the paper will exceed permissible length. Any excess explanatory items should be submitted separately in case of overlength
   （这个确实，不过就你们这比水平，我还真怕写少了，你们看不懂！）
6. There was no mention and discussion on 2 highly relevant references from the literature which talk about the utilization of correlations. The authors must include these and compare their work in contrast to these publications as well:
   a) Xifilidis, Theofanis, and Kostas E. Psannis. “Correlation-based wireless sensor networks performance: The compressed sensing paradigm.” Cluster Computing (2022): 1-17.
   b) Akyildiz, I. F., Vuran, M. C., & Akan, O. B. (2004, March). On exploiting spatial and temporal correlation in wireless sensor networks. In Proceedings of WiOpt (Vol. 4, pp. 71-80).
   （这两个是高度相关的工作？笑死我了，我是做定位的，这两个是做什么的你不看看吗？这叫高度相关？）
7. Authors are requested to also compare or comment on the energy efficiency of the proposed technique when comparing against other methods for various applications.（真是泰库拉，第一次见到比能量和效率的，在定位的算法论文里面，那定位里面的著名论文landmark也使用了kNN聚类的机器学习算法也没有见比较啊。）

Additional Questions:

1. Which category describes this manuscript?: Research/Technology

2. How relevant is this manuscript to the readers of this periodical? Please explain under Public Comments below.: Relevant

3. Please explain how this manuscript advances this field of research and/or contributes something new to the literature.: The manuscript addresses the limitation of existing wireless sensing methods by utilizing signal correlation in channel, time, and space. The experimental results indicate the effectiveness of the proposed framework in improving the accuracy and robustness of wireless sensing applications, and therefore contributes a new approach to the literature.

4. Is the manuscript technically sound? Please explain under Public Comments below.: Appears to be - but didn’t check completely

5. Are the title, abstract, and keywords appropriate? Please explain under Public Comments below.: Yes

6. Does the manuscript contain sufficient and appropriate references? Please explain under Public Comments below.: Important references are missing; more references are needed

If you are suggesting additional references they must be entered in the text box provided. All suggestions must include full bibliographic information plus a DOI.

If you are not suggesting any references, please type NA.: 1) Xifilidis, Theofanis, and Kostas E. Psannis. “Correlation-based wireless sensor networks performance: The compressed sensing paradigm.” Cluster Computing (2022): 1-17.

2. Akyildiz, I. F., Vuran, M. C., & Akan, O. B. (2004, March). On exploiting spatial and temporal correlation in wireless sensor networks. In Proceedings of WiOpt (Vol. 4, pp. 71-80).

3. Does the introduction state the objectives of the manuscript in terms that encourage the reader to read on? Please explain under Public Comments below.: Yes

4. How would you rate the organization of the manuscript? Is it focused? Is the length appropriate for the topic? Please explain under Public Comments below.: Could be improved

5. Please rate the readability of this manuscript. Please explain your rating under Public Comments below.: Easy to read

6. Should the supplemental material be included? (Click on the Supplementary Files icon to view files): Yes, as part of the digital library for this submission if accepted

7. If yes to 6, should it be accepted: As is

8. If this manuscript is an extended version of a conference publication, does it offer substantive novel contributions beyond those of the previously published work(s)- i.e. expansion of key ideas, examples, elaborations etc. New results are not required: Not applicable

Please rate the manuscript. Please explain under Public Comments below.: Fair

Reviewer: 2

Recommendation: Reject

Comments:
In this manuscript, the authors study an interesting problem, i.e., learning correlation among wireless signals via graph neural network. They propose a Signal Correlation Learning (SCL) framework which represents three types of correlation among wireless sensors and apply this framework to different learning tasks as well as different wireless sensors. However, the paper does have the following issues.

First, the contribution of the paper is not that significant. Some main ideas such as channel attention, graph construction, or graph attention have already appeared in existing works. Although the authors made some change to them, the explanation of the change is not clear and the evaluation is insufficient. For example:

1. At the end of section 3.4, how do the authors get an appropriate value of k according to the properties of small world? How sensitive is the framework w.r.t k?（牛逼啊！TOPK在时间序列为图里面是多么常见的算法在你这儿就不行了？至于K怎么取值？我建议你和之前那个审稿人battle一下，他觉得我写的太长了，所以我就不把小世界网络建图的过程详细写了。至于K取值影响的效果？消融实验你是看不到吗？）
2. In section 3.4, can the constructed directional graph contain bidirectional edges? If no, how to represent two sensors with similar noise and are highly correlated with each other? If yes, how is such a graph transformed to a forest as shown in Fig. 16? By the way, it would be better if the authors can compare the generated graph with the real position of the sensors when doing the interpretation.（双向边？都他妈在文章中说了是有向动态图了，你问我有没有双向边？你别告诉我你不知道什么叫有向图？至于怎么表示相似传感器？那边权重和TOPK的连边是干嘛呢？我们为什么要做这个事情？不就是在做你说的事情了吗？都说了TOPK选取高维空间中前K个距离最近的作为连边，距离作为权值。你咋还问呢？而且这么理直气壮？至于传感器和真实位置？都已经使用图这种非欧式空间的结构了扯什么物理位置？）

Second, some settings of the experiments are questionable. For anomaly detection, it is unclear what kind of anomalies are added into the data. For localization, it is unclear how big the test site is. From Fig. 8, it seems that the test site is quite small, far different from the real-world indoor localization scenario. Without knowing such information, it is difficult to judge whether the proposed framework can work in practice.（异常检测场景作为一个case study，是说除了定位我们还能做异常检测。至于详细的说明，兄弟RFID克隆攻击？这个说明还不够详细的吗？你可是TMC的审稿人啊！！RFID中的伪造标签不知道？你就算不知道我后面不是给了参考文献？你去看一下啊！你要是觉得这东西我应该在论文里面写出来，那我建议你也去和第一个审稿人打一架，他说我写太多了，14页了。至于场景小，那TM NFC的全称是什么，不就是进场通信吗？RFID的通信距离有多远？）

Moreover, some related works about the application of graph neural network in localization or anomaly detection are not mentioned in the paper, such as:

Guan, Siwei, Binjie Zhao, Zhekang Dong, Mingyu Gao, and Zhiwei He. “GTAD: Graph and Temporal Neural Network for Multivariate Time Series Anomaly Detection.” Entropy 24, no. 6 (2022): 759.（这篇论文我要是引了我把你眼珠子挖出来可以不，打个赌不？）

Zheng, Han, Yan Zhang, Lan Zhang, Hao Xia, Shaojie Bai, Guobin Shen, Tian He, and Xiangyang Li. “Grafin: An applicable graph-based fingerprinting approach for robust indoor localization.” In 2021 IEEE 27th International Conference on Parallel and Distributed Systems (ICPADS), pp. 747-754. IEEE, 2021.

Luo, Xuanshu, and Nirvana Meratnia. “A Geometric Deep Learning Framework for Accurate Indoor Localization.” In 2022 IEEE 12th International Conference on Indoor Positioning and Indoor Navigation (IPIN), pp. 1-8. IEEE, 2022.
（以上两篇论文没有引述其实是一个很麻烦的事情，因为我这个论文最早完工的时候是前年，当时我发现GNN可以应用于无线感知的各个领域而且应该可以取得一个不错的效果，其实这是一个很简单的思想，无线感知经常使用参考信号，会根据已知信号和未知信号之间的关系进行推理，那很自然的想法是我们能不能用一些方法去自动挖掘出信号之间的关系呢？那很容易想到GNN，之后做出来的时候投了NSDI被拒了，后面会放上NSDI的审稿意见。然后拖了大半年改投TMC直到现在才出审稿意见。所以这么长的时间跨度肯定也有人做出类似的工作，所以没有办法引用哪些论文了。）
Also, some typical deep graph neural networks such as the graph attention networks [14] are not included in the baseline.
（哈哈哈哈哈，我的baseline里面有两篇时间序列为图的经典论文一个是图差分网络GDN、和MTAD。这两个都使用了GAT作为骨干网络，怎么能叫没有使用GAT作为baseline呢？并且我在baseline介绍里面已经说的很清楚了。你是看不懂还是看不到吗？）
Finally, there are quite a lot typos in the paper, which may confuse the readers. For instance:

1. In the last paragraph in section 2.2. It is \sum{(x,e)}, not e that is called the Fourier coefficient.
2. In formula 9, “-1” should be the superscript instead of the subscript.
3. In formula 12, “U\hat{f}” seems to be “U\hat{A}”.
4. In the second paragraph of section 5.1, “dropper shift” -> “doppler shift”.
5. In the second paragraph of section 6, “nodes B and F whose in-degree is 0” -> “nodes A and F whose in-degree is 0”.
   （这就是我说的中间拖了大半年，我之前写的时候没有这些问题，某个人改完之后这些问题又来了。无语！别问我为什么别人改完之后自己不再看一遍？）

Additional Questions:

1. Which category describes this manuscript?: Research/Technology

2. How relevant is this manuscript to the readers of this periodical? Please explain under Public Comments below.: Relevant

3. Please explain how this manuscript advances this field of research and/or contributes something new to the literature.: This manuscript introduces a signal correlation learning framework that represents the signal and learns the relationship of wireless sensors via a graph neural network. This framework can be applied to different sensing problems including localization, anomaly detection, and human activity recognition with different wireless sensors including RFID and Bluetooth.

4. Is the manuscript technically sound? Please explain under Public Comments below.: Appears to be - but didn’t check completely

5. Are the title, abstract, and keywords appropriate? Please explain under Public Comments below.: Yes

6. Does the manuscript contain sufficient and appropriate references? Please explain under Public Comments below.: Important references are missing; more references are needed

If you are suggesting additional references they must be entered in the text box provided. All suggestions must include full bibliographic information plus a DOI.

If you are not suggesting any references, please type NA.: Guan, Siwei, Binjie Zhao, Zhekang Dong, Mingyu Gao, and Zhiwei He. “GTAD: Graph and Temporal Neural Network for Multivariate Time Series Anomaly Detection.” Entropy 24, no. 6 (2022): 759.

Zheng, Han, Yan Zhang, Lan Zhang, Hao Xia, Shaojie Bai, Guobin Shen, Tian He, and Xiangyang Li. “Grafin: An applicable graph-based fingerprinting approach for robust indoor localization.” In 2021 IEEE 27th International Conference on Parallel and Distributed Systems (ICPADS), pp. 747-754. IEEE, 2021.

Luo, Xuanshu, and Nirvana Meratnia. “A Geometric Deep Learning Framework for Accurate Indoor Localization.” In 2022 IEEE 12th International Conference on Indoor Positioning and Indoor Navigation (IPIN), pp. 1-8. IEEE, 2022.

3. Does the introduction state the objectives of the manuscript in terms that encourage the reader to read on? Please explain under Public Comments below.: Yes

4. How would you rate the organization of the manuscript? Is it focused? Is the length appropriate for the topic? Please explain under Public Comments below.: Could be improved

5. Please rate the readability of this manuscript. Please explain your rating under Public Comments below.: Readable - but requires some effort to understand

6. Should the supplemental material be included? (Click on the Supplementary Files icon to view files): Does not apply, no supplementary files included

7. If yes to 6, should it be accepted:

8. If this manuscript is an extended version of a conference publication, does it offer substantive novel contributions beyond those of the previously published work(s)- i.e. expansion of key ideas, examples, elaborations etc. New results are not required: Not applicable

Please rate the manuscript. Please explain under Public Comments below.: Fair

Reviewer: 3

Recommendation: Reject

Comments:
The paper presents and discusses a Signal Correlation Learning (SCL) framework adopted to define a graph that represents the signal correlation of multiple wireless sensors. My comments after reading the paper are as follows:

• The authors should revise their paper concerning formatting e.g., missing spaces, etc（什么地方嘛？像第二个审稿人那样给个位置啊？）
• The authors should explain better what is the concept of signal correlation especially the spatial and temporal correlation. Also, they should explain what are the hidden parameters of the signal that could be detected through the use of deep learning frameworks.（你是在这里面跟我要可解释性吗？我解释泥马勒戈壁我解释，我要是能解释了还投泥马勒戈壁的TMC，直接投natural不比你这垃圾期刊强太多了。）
• What is the meaning of channel-, spatial-, and temporal-correlation features? Are these features of a signal? Is there any distribution behind the spatial placement of sensors or the selected channel?（你的意思是通道、空间和时间相关特征的含义是什么？这些是信号的特征吗？传感器或所选通道的空间位置后面是否有分布？是这样吗？来你告诉我视频的通道、空间和时间的相关特征是什么？我发现了一个问题就TMC这群币审稿人很难理解说GNN或者多元时间序列里面的spatial的这个意思，他们不知道通过相似度计算得到的图的spatial和物理空间中space的区别。这也就罢了，你一个TMC的审稿人不知道通过时间采样得到一段信号，不理解这个是什么意思吗？而且更离谱居然不了解什么是通道，我认为OFDM这种东西属于是无线的基础了吧，N个正交的子载波等于N个通道，这个算是无线的基础了吧，这玩意不知道吗？）
• The authors do not describe the motivation behind the pre-processing phase. Various questions arise like why we need the interpolation or the normalization steps.（来我把你的问题翻译一些：作者没有描述预处理阶段背后的动机。出现了各种各样的问题，比如我们为什么需要插值或归一化步骤。牛逼啊！我第一次见到数据预处理方面居然是要详细解释的，但是问题是我也解释了，插值是为了让数据长度对齐，归一化是为放缩消除指标之间的量纲影响，你是看不懂吗？）
• Till page 6, the authors do not explain what the correlation is. Is it the statistical measure upon the signals, is it the distance between sensors, etc? A number of metrics are mentioned in the paper, however, there is not any clear indication about the specific metric that is adopted in the model.（来来来，翻译一下这个牛逼的问题，在第6页，作者没有解释相关性是什么。它是对信号的统计测量吗，是传感器之间的距离吗，等等？论文中提到了许多度量，但没有任何关于模型中采用的具体度量的明确指示。我论文里面写的是计算经过通道注意力机制的数据在高维空间中的KL散度。这一句话完全可以回答你所有疑惑了吧，咋的，你不明白？就是我说的我发现了一个问题就TMC这群币审稿人很难理解说GNN或者多元时间序列里面的spatial的这个意思，他们不知道通过相似度计算得到的图的spatial和物理空间中space的区别。）
• The authors should explain better how the temporal correlation is depicted by the envisioned graph. Is it one graph or multiple graphs? How the dimension of time is incorporated in the graph?
  （我就觉得这群审稿人怎么这么普通却又这么自信呢？我觉得一个有向动态图外加动态图的数学表示已经够了吧，没必要介绍动态图的概念了吧。你要是说我需要介绍动态图是什么，我建议你去和第一个审稿人打一架。因为长度已经超标了。）
• The adoption of deep learning models is not adequately excused. I know that the use of deep learning models is a trend nowadays, however, the authors should explain why they rely on the LSTM architecture. （牛逼啊又是和副主编一样的问题，你们问题能不能换点花样？那么多动态图用LSTM+GNN的算法在我的参考文献里面你是看不着吗？至于评估，我知道你们审稿人从来不看消融实验。也许应该我直接在论文里面一同吹牛逼而不做实验，那样可能符合TMC的评估。）
• Additionally, only limited information is provided for the training dataset and how this dataset of 6000 data points could be the appropriate one to avoid overfitting. Additionally, the authors should explain what happens when the dynamics of the system significantly change and a re-training may be required (?).（真的，审稿人我建议你可以去跑一下pyg里面的demo，真的，你看一下空手道俱乐部里面的数据集里面才多少个点，为什么没有过拟合？你去了解一下什么是半监督的节点分类问题在来这里面叫可以不？或者你不去了解而是看看我论文里面的问题定义部分，我在里面清楚地写了，我将定位问题转化为了图上地节点回归问题。啊，我知道你们TMC地审稿人一向是不看地。至于环境信号发生显著变化？那当然要重新训练了，我要能整出一个这么鲁棒地模型，我直接投natural不好？投什么TMC）
• The authors should elaborate more on the overhead that the training process of neural network adds into the system. For instance, if we want to depict the temporal correlation of vectors or streams why another simple statistical process is not appropriate and we need the deep learning model?（训练开销，为什么所有地审稿人都这么问？问地我以为我在投NSDI呢！没想到TMC也是一个系统类地期刊啊！至于为什么不使用传统地统计学模型，我不会告诉你空间谱分解和GNN本身就有着极大的关系。为什么我不告诉你呢？因为实验部分已经证明了传统算法不如GNN+LSTM+通道注意力机制了。你就是不看啊！）
• The use of the output of the system (fig. 6) is not clearly presented. What decisions are supported and how those decisions will increase the performance of the system?（我就没看懂你这个问题的意思，是说这个系统支持哪几种决策？首先我们做的不是一个系统，而是引入了一种处理定位的新思想。至于决策？我理解的是是不是这个系统支持哪几种case study呢？那不是写的很清楚吗？主要适用于定位，也可以应用于异常检测、行为识别。不知道你在说什么。）

Additional Questions:

1. Which category describes this manuscript?: Research/Technology

2. How relevant is this manuscript to the readers of this periodical? Please explain under Public Comments below.: Relevant

3. Please explain how this manuscript advances this field of research and/or contributes something new to the literature.: There are questions about the novelty of the approach.

4. Is the manuscript technically sound? Please explain under Public Comments below.: Appears to be - but didn’t check completely

5. Are the title, abstract, and keywords appropriate? Please explain under Public Comments below.: Yes

6. Does the manuscript contain sufficient and appropriate references? Please explain under Public Comments below.: References are sufficient and appropriate

If you are suggesting additional references they must be entered in the text box provided. All suggestions must include full bibliographic information plus a DOI.

If you are not suggesting any references, please type NA.: NA

3. Does the introduction state the objectives of the manuscript in terms that encourage the reader to read on? Please explain under Public Comments below.: Could be improved

4. How would you rate the organization of the manuscript? Is it focused? Is the length appropriate for the topic? Please explain under Public Comments below.: Could be improved

5. Please rate the readability of this manuscript. Please explain your rating under Public Comments below.: Easy to read

6. Should the supplemental material be included? (Click on the Supplementary Files icon to view files): Does not apply, no supplementary files included

7. If yes to 6, should it be accepted:

8. If this manuscript is an extended version of a conference publication, does it offer substantive novel contributions beyond those of the previously published work(s)- i.e. expansion of key ideas, examples, elaborations etc. New results are not required: Not applicable

Please rate the manuscript. Please explain under Public Comments below.: Fair
综上这次TMC投稿给我的感觉非常的差，先是审稿申了这么长时间，其次就这审稿意见甚至比不上chatpaper自动生成的。一个图神经网络的论文，三个意见包括副主编的意见居然和图神经网络没有任何关系。真的审成这样也是没谁了。真的我看网上喷好多人说 infocom、AAAI、IJCAI水，就是因为审稿人水平差，利益相关者。但是我这次一看TMC这种臭水平的垃圾审稿人，真的别喷infocom AAAI了，我现在才知道原来期刊的水比会议深太多了。对了我之前投过NSDI，以下是NSDI的审稿意见：
Overall merit
1.
Reject (I will argue to reject this paper)

Reviewer expertise
1.
No familiarity

Paper summary
The paper proposes to use signals–collected by one device about multiple other devices–to construct a signal correlation learning framework which can then be used to solve multiple sensing tasks.

Comments for authors
The tasks you describe, e.g., anomaly detection, localization, and human activity recognition, are very challenging and require novel solutions. The idea of taking advantage of the correlated signals is also excellent. However, this reviewer does not think this is the first time someone is proposing to take advantage of correlated signals. For example, Emerald – a company that is using wireless signals to infer health signals – is using correlated wireless data to do so. Moreover, the correlation of signals is used to improve wireless data communication etc.

As such this reviewer is confused by some of the claims of the authors. Overall, this paper seems to be much more focused on explaining the ML setup of the system rather than the application side. As such it is unclear why NSDI is the best conference for such a paper.

Another aspect that the paper does not deal with is that wireless signals are constantly changing, e.g., because a device is moved, because of temperature, because of a human moving, because of another device that is active (e.g., the microwave) etc. In general, the data sets that you are using are rather small and do not span long time periods. You claim that your system is self-adaptive but you never challenge it.

Indeed, before you can even discuss your SCL framework you should first state what kind of data you are looking at. Next, you should explain your assumptions about the problems you want to solve as well as the data. Next, you should explain what the components of your model are, e.g., what is the meaning of a node and a graph, and how does time play into this. For example, up to page 4, you talk about a graph but never explained what your nodes are…

You also talk about a signal correlation graph but you do not talk about how you measure correlation. In particular, you do not talk about how you explicitly caption the three different forms of correlation. Also, what is the complexity of handling the correlation?

Regarding your results, you claim a factor of up to 2.5 improvements over the baseline… But the relevant number is the improvement over the best alternative. This seems to be significantly smaller.

Review #503B
Overall merit
1.
Reject (I will argue to reject this paper)

Reviewer expertise
4.
Expert

Paper summary
The paper proposes a Signal Correlation Learning (SCL) framework, a data-driven method based on Graph Neural Network (GNN) to find signal correlations in various sensing tasks. SCL constructs a direct graph to represent the channel-, spatial-, and temporal- signal correlations of wireless sensors. The framework preprocessed data, assigns weight, constructs a graph using KL-based method, then correlates features. The method is evaluated on RFID/bluetooth localization, anomaly detection, and human sensing to demonstrate its performance.

Comments for authors
It’s interesting to see work that tries to build general approaches for wireless sensing. However, it was difficult for me to understand the practical problem that this paper is trying to solve. Is the goal to come up with more elegant solutions to existing problems? Or is the goal to outperform prior systems? It seems that the paper is trying to do the former but is pitching the latter, yet I believe it underperforms in comparison to prior art.

The paper focuses on 3 tasks, and says it outperforms prior art by 1.5-2.5x. I’m not sure that is true for any of the tasks:

If you take RFID localization, the paper achieves 20cm accuracy (Fig. 9b), but state-of-the-art systems achieve 1-cm of accuracy. So, the paper underperforms by 20x. In comparison to an old baseline like Landmarc from the early 2000’s, it achieves better performance, but that’s not an acceptable baseline
If you take activity recognition, the paper only studies 3 activities. State-of-the-art systems can either precisely recover trajectories/movements (e.g., work from Dina Katabi’s group) or can classify at least a dozen movements. So, I’m not sure the paper outperforms there either.
Finally, the anomaly detection task was unclear. The paper states that they “invade the data and replace them with different signals to simulate the anomaly attacking” - what does this mean? Is this a problem to solve?
Overall, exploring more elegant or simpler solutions to exiting problems/tasks might be intellectually interesting, but it’s not clear to me that this paper does that either, and it seems to underperform in comparison to prior art.

Few other detailed comments:

The paper states that SCL needs multiple sensors to coexist to work. I wonder how many sensors do we minimally need for the SCL to work? The paper doesn’t mention this and only states “several” E41C Impinj tags, instead of an actual number.
What’s the computation complexity for the method? SCL is currently using a high-performance GPU server to run the inference. I could imagine that many applications would like to run the wireless sensing application on edge devices. Therefore, a brief understanding of the computing complexity, or the latency of the whole inference task will give us a better understanding of the framework.
From the evaluation of SCL on RFID and Bluetooth systems on different sensing tasks, I didn’t see much difference between them. I wonder why we need three different channels (RSSI, Phase, Doppler) for RFID while Bluetooth only needs RSSI and achieve almost the same accuracy.
Some typos:

There are some typos in the paper that wrote SLC instead of SCL.
On page two, in this sentence: “In this paper, we perform the weighted average operation on the the feature”, it uses two “the”.
Review #503C
Overall merit
2.
Weak reject (I think it should be rejected, but I am fine if others want to accept.)

Reviewer expertise
2.
Some familiarity

Paper summary
This paper presents a correlation learning mechanism to improve sensing accuracy in IoT deployments. The authors evaluate their framework in a testbed and find 1.5-2.5x improvement in sensing accuracy relative to systems employing existing techniques.

Comments for authors
Thank you for submitting to NSDI.

While the claimed improvements and novelty sound significant, I have a primary concern over the presentation:

for a systems conference, there is insufficient overview of the techniques employed and their novelty. The paper unfortunately goes straight into the math without any description of the protocol in operation.
related, beyond the insight of using signal correlation, what is hard about the work? Is this a straightforward adoption of signal correlation? Or is their insight into the approach as well.
while there is some discussion of related work, I would have liked to have seen more on the employment of signal correlation from related areas.
this is a problem beyond this paper, but the evaluation is fairly ad hoc. While detailed, the challenge with IoT is that the deployments are rather heterogeneous/varied. Does this work solve a practical problem in the field? Or is it speculating about challenges that might come along in the future?
看到没有，不对比就没有差距。而且咱也不是那种接受不了别人的批评那种，NSDI虽然给我拒了但是我心服口服啊！人家审稿人看不懂的地方就直说我看不懂。然后人家就比较人家能看懂的地方是否符合会议的要求。最终认为不符合给拒绝了。这样我也认了。这才叫专业的审稿人啊！谦虚、客观、有礼貌、就事论事、没有攻击性。这样一比那帮TMC的臭鱼烂虾怎么有脸活下去呢？