Convolutional Long Short-Term Memory Two-Dimensional Bidirectional Graph Convolutional Network for Taxi Demand Prediction
Round 1
Reviewer 1 Report
This paper introduces a framework named CTBGCN for taxi demand prediction at the level of OD pairs. Overall, the paper is well written and technically sound, with a prediction accuracy higher than several existing methods. I recommend the publication of this paper in the journal. I have only one suggestion as follows: I am curious but unsure if the authors could further look into some detailed results regarding the prediction accuracy among the OD pairs of different regions.
Specifically, the distribution of taxi demand across regions in a city is quite heterogenous, with only a few hub regions possessing larger demands. The authors’ proposed method shows superiority over existing methods, but we do not know whether such high accuracy can be equally achieved for the OD pairs of all the regions in a city or is only for regions of larger taxi demands. Therefore, it would be even better if the authors could divide regions into several types according to their taxi demands and show some detailed results of the prediction accuracy for different types of regions.
Author Response
Please see the attachment.
Author Response File: 
Author Response.pdf
Reviewer 2 Report
This paper provides results and findings for applying Conv-LSTM to estimate taxis' OD distribution based on a dataset obtained in New York. The followings are some suggestions to improve this paper:
1) The authors have never spelled out "Conv-LSTM" and this should be amended and explanations provided.
2) In line 91, Related works should not be "recommended" but "reviewed". In line 110, it should be made clearer any shortcomings of the traditional methods that have driven the development of the particular method proposed in the paper. "machine learning" should not be considered as a traditional method though as presented in line 101.
3) A number of abbreviations were used in this paper, such as LSTM, ...etc. The authors should make sure that the full spelling of the abbreviations involved should be mentioned in their first occurrences in the paper.
4) In line 119 and other similar occurrences, the authors should avoid using the name of abbreviated terms as the subject, e.g., "FCL-Net [5] proposed ...", instead the authors should amend the same statement in the following manner: "It was proposed in [5] that the FCL-Net was used to fuse ..."
5) The authors should check carefully that an article should be included for singular term, e.g., in line 144, "Graph structure..." should be "A graph structure..." All other similar occurrences in this manner should be checked and amended.
6) In line 153, it's not sure about the meaning of the sentence. The authors may wish to state in a more meaningful way.
7) In line 151, "slightly overused ..." what does that mean?
8) In line 183, the authors should explain why as per what'e mentioned, "as much work does"?
9) Equation (2) contains a number of symbols and mathematical operations that requires the authors to present clearer explanation on how they represent the process involved in Conv-LSTM. The same comment is also given to equations (3) and (4).
10) In line 197, "We use two chapters to ..." Not sure which two chapters the authors were intending to refer to?
11) In line 230, please explain why the authors thought the change is "creative"?
12) In line 242, the same dataset has been used in many other research and their associated published papers. The authors should provide more information on the dataset and the referenced sources. Moreover, the authors should also mentioned whether any similar works have been produced and whether any comparable results can be discussed in this paper.
13) In line 246, "between 0 and 030" should be "between 00:00 and 00:30)"
14) In Tables 1 and 2, "CTBGCN (ours)" actually meant "Conv-LSTM". As a general rule of thumb, it's always better to use a third person's tense to write in the paper. Avoid too many "We" or "ours", as they may reduce the objectiveness of the discussions in the paper.
15) The authors presented the results of the goodness of fitting the data, but it doesn't seem the predicability of their proposed models, which seems to be the main theme of the paper.
16) As indicated in line 323, the improvements were only slightly. The authors should comment what other benefits they observe with their proposed model. This is important, as otherwise, this paper simply reported an academic exercise to try a different method.
17) The conclusion is too brief and lacking substance, with many questions unanswered, and some of the claims made by the authors were not made noticeable with supports of results. For example, lines 293 to 303, the authors should be more specific with results found to substantiate their associated claims of the superbness of their proposed model.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Reviewer 3 Report
Taxi demand prediction has gained more and more research interest nowadays, thus, the authors adopt a new spatial-temporal model named Conv-LSTM Two-dimension Bidirectional Graph Convolutional Network (CTBGCN) to predict taxi demand in the manuscript. The structure of manuscript is easy to follow, howerver, I recommend major revision which need to be resolved.
1. Readers may find it difficult to understand how CTBGCN operate; therefore, some pseudocode should be provided.
2. How do the authors convert the dataset's data into CTBGCN's input data? It is necessary to provide a brief example.
3. To illustrate the difference between expected outcomes and actual results, it is preferable to offer a comparison figure in the results section.
4. Figure 4 should include a color legend that explains what the various shades represent.
5. A more detailed conclusion, including limitations and future improvements to the proposed method, should be provided.
6. There are numerous typos in the current manuscript; the authors should review and correct them.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Round 2
Reviewer 3 Report
The authors addressed most of my concerns, but some questions about CTBGCN's pseudo-code remain.
(1) Why does input x (n, N, H, W) and output y (N, H, W) have different shape?
(2) Does the symbol N has the same meaning in section 3.1? It is simple to become perplexed.
(3) Why does C2 equal to 64?
(4) What do Ao, Ad, Eo and Ed mean?
(5) Why uses MLP twice in line 8?
Author Response
Please see the attachment.
