Adaptation and Learning to Learn (ALL): An Integrated Approach for Small-Sample Parking Occupancy Prediction
Round 1
Reviewer 1 Report
The article from a methodological point of view meets the expectations of a scientific article. The authors establish clear categories of analysis and comprehensively explain the procedure used for both data collection and learning models. The methodology is also very clear and well worked out from a practical point of view.
However, it is said that it is impossible to accurately estimate the minimum amount of data required for an AI project. Accurate estimates take many other factors into account, one of which is analyzing what we can do to enhance our data set before exploring technical solutions. It may sound obvious, but before you get started with AI, try to get as much data as possible by developing external and internal tools with data collection in mind. Moreover, there are some general approaches that can help you build predictive models from small data sets, such as naive Bayes. They also say that starting with a simple heuristic is effective. Heuristics have several advantages. It does not require significant amounts of data and can be generated by intuition or domain knowledge. They have a high level of interpretability. Moreover, it does not suffer from problems or complications with data skew or sensitivity of the model itself. Therefore, in a chapter such as Discussion, please state the reason why you did not use a method that enhances the data itself, does not include a heuristic method, and uses a method that only considers the learning model.
Author Response
Dear Reviewers,
We would like to express our sincere appreciation for your careful reading and valuable comments to improve this paper. We have addressed all issues raised by the reviewers.
Please see the attachment.
Author Response File: 
Author Response.pdf
Reviewer 2 Report
In this manuscript, the authors have proposed an Adaptation and Learning to Learn (ALL) approach by adopting the capability of advanced deep learning and federated learning to address challenges such as the insufficient amount of the data in Parking occupancy prediction (POP). The experiments show the effectiveness of the proposed framework.
Suggestions that would improve the quality of the paper.
The dataset used in this study is very old. Authors are suggested to conduct experiments on the latest data.
Provide an architecture diagram of the proposed method.
How did you apply Deep Learning? Please explain it.
Author Response
Dear Reviewer,
We would like to express our sincere appreciation for your careful reading and valuable comments to improve this paper. We have provided a point-by-point response to your comments.
Please see the attachment.
Author Response File: Author Response.pdf
Reviewer 3 Report
The paper presents a knowledge transfer approach that integrates two ideas, 1) using the asynchronous advantage actor-critic Reinforcement Learning Technique, an auto-selector module is developed to automatically select data-scarce parks as supporting sources to enable knowledge adaptation in model training, and 2) through federated meta-learning of selected supporting sources, a meta-learner module is developed, which can train a high-performance local prediction model in a collaborative and privacy-preserving manner. The manuscript is generally structured and written well.
Though I understand this is not a theoretical paper, my main concern has to do with the contributions of the paper. It is clear that the authors do not propose any theoretical novelties in the problem of learning. The main contribution, the way that I understand it, is to use the asynchronous advantage actor-critic Reinforcement Learning and federated meta-learning techniques to handle data shortage in data-poor parking areas. However, although each technique employed is known, some of the ways that the authors combine those techniques are not adequately motivated.
It is well-known that a single critic network setup is computationally more efficient than an actor-critic setup (see for examples: 10.1109/JAS.2017.7510784 or 10.1109/CDC.2010.5717676). So, my question is, why do authors choose the actor-critic setup? What will be the challenges of implementing the proposed framework using a single critic network instead of actor-critic setup? Some explanation would be helpful for the readers.
The paper requires careful proofreading. There are a few typos that should be fixed. For example,
"to generates synthetic samples " -> "to generate synthetic samples "
"and combine the multiple-graph convolutional"->"and combines the multiple-graph convolutional"
"have heavy dependence on availability of other data "->"have heavy dependence on the availability of other data "
"methods built on the idea of transfering knowledge"->"methods built on the idea of transferring knowledge"
"to implement the transfer learning based "->"to implement transfer learning based "
"Considering the large amount of parking lots in cities"->"Considering a large amount of parking lots in cities"
" (i.e. parking lots) through exchanging"->" (i.e., parking lots) by exchanging"
Despite the stated comments, I think the paper is worthy of being published.
Author Response
Dear Reviewer,
We would like to express our sincere appreciation for your careful reading and valuable comments to improve this paper. We have provided a point-by-point response to your comments.
Please see the attachment.
Author Response File: Author Response.pdf
