Listening to the City, Attentively: A Spatio-Temporal Attention-Boosted Autoencoder for the Short-Term Flow Prediction Problem

Round 1

Reviewer 1 Report

This paper proposed an autoencoder-based neural network for the flow prediction problem. 

 

Strength:

1. Important problem. This paper deals with the flow prediction problem, which is very important for urban traffic-related tasks.

2. Clear presentation. This paper is well-structured, which informative figures and tables.

3. Easy-to-reproduce experiments. This paper uses three open-sourced datasets for experiments and their codes are open-accessed.

 

 

Weakness:

1. There is a bunch of related work that could be included in the Related Work part or even as baselines.  Authors please refer to LibCity (https://github.com/LibCity/Bigscity-LibCity) for a more comprehensive introduction to related work and explain how you chose the baselines.

2. It's not clear why CMU is needed, especially for the flow prediction problem. It is suggested to explain why CMU/MU is better suited for your task with intuition (not just with experimental results).

 

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

The motivation of this paper is convincing, and I believe that this paper addresses an important research problem in mobility data analysis. Here are some strong points of this paper: 1) the presentation of this paper is good. The paper is easy to follow and should address the interests of mobility researchers or transportation researchers. 2) the experiments of this paper cover a wide range of comparisons. The authors of this paper have listed a long list of baseline methods to compare with their proposed model, which makes the experiment section convincing and solid. However, this paper also has some weaknesses:

1.       Lack of technical depth. The authors stated that they propose a novel deep learning network with a multi-head attention mechanism that captures and exploits spatial and temporal patterns in mobility data. However, the design of the model is incremental, and in the discussion of the attention mechanisms, the authors do not provide detailed calculations on the attention weights. Questions such as what is defined as the keys, values, and queries are not being discussed.

2.       Topic mismatch. The journal is for algorithms, but this paper proposes a deep learning model, and the actual solutions/algorithms of this deep learning are provided by third-party deep learning frameworks such as TensorFlow or PyTorch. And I don’t see any algorithmic innovations in this paper. However, this point should be judged by other editors as well.

3.       This paper contains numerous typos and grammatical errors, and careful proofreading is highly recommended. To name a few:

n  In line 65: “the study demonstrate” à “the study demonstrates”

n  In line 84: “Artifical” à “Artificial”

n  In line 88: “Others approaches” à “Other approaches”

n  In line 90: “statistics models shows” à “statistics models show”

n  In line 109: “which exploits” à “which exploit”

n  In line 307: “devised to to” à “devised to”

 

There are more grammatical errors in the case study section, and I don’t enumerate every one of them here. 

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

This paper solves the traffic flow prediction problem. This problem is important for cities and businesses in providing predictions to serve their short-term and long-term needs. The authors propose a new deep learning structure that takes advantage of a novel set of deep learning technologies. Each of these proposed technologies address a specific challenge in this specific problem, guided by the existing knowledge on displacement dynamics. The authors demonstrate that their method outperforms the many existing methods in the literature.

The paper is of acceptable quality, with good amount technical complexity to address the complexities of the problem. However, my main concern is novelty. My specific concerns are as follows:

1. This is a crowded field and previous deep learning methods have already addressed the issue of dependence between space and time. For instance, any method that uses LSTM and convolution layers together, do so to address this dependence. Therefore, this cannot be claimed as the novelty of this work, because it has been done before.

2. In the introduction, the authors mention two ways the existing patterns are disrupted: COVID-19 and new means of transportation, such as bike-sharing, ride-sharing, e-scooters, etc. It gives the impression to the reader that the novelty of the current work is to propose a method that is designed to be robust against such emerging disruptions. However, I am having a hard time connecting these remarks to a specific contribution or design decision in the following sections.

3. In the Related Work section, the authors must explicitly mention their novelty and differences from the existing art. Just reviewing the literature without connecting it to the novelties of the current work is not sufficient to justify why this work is necessary.

4. Two of the instances where the authors explicitly mention how they differ from others are Page 5, Line 199 and Page 7, Line 229. These mention design decisions that differ from the literature. If these are the claimed novelty of the work. they should be translated to non-technical language and be included in Related Work and Introduction sections.

5. Another brief mention of the novelty is in Page 2, Line 60, which also mentions a design decision, without an elaboration on what it means in the context of this problem.

In conclusion, I believe even though the paper is well-written and the solution is technically sound, the novelty of the work is not sufficient at its present form. The paper's novelty is limited to a few design decisions in its solution, with no elaboration on what those decisions mean in the context of this problem and what limitations and/or challenges they address that has not been addressed before. Unless the authors can provide such an elaboration, the publication of this paper cannot be justified, because the novelty is unclear and the improvements are minor.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 2 Report

I think the authors have addressed all my concerns. 

Reviewer 3 Report

After extensive revision, the authors have fully clarified the novelty of the work, which was my main concern. As I mentioned in the previous round of review, the paper is technically solid and now that the novelty concern is addressed, its quality has improved significantly. Even though the contributions are still limited to design and accuracy improvements, the design decisions are innovative enough and the publication of this paper would be beneficial to the research community. I recommend accept.