DynST: Large-Scale Spatial-Temporal Dataset for Transferable Traffic Forecasting with Dynamic Road Networks

• The paper appears to require daily generation of the dynamic road network topology using the tree-based adjacency matrix generation algorithm. The efficiency of this process remains unclear. Additionally, since the topologies undergo minimal changes between consecutive days, and substantial information is shared across these days, it raises the question of whether specialized algorithms are available to accelerate this topology generation.
• The authors present performance results only for two districts, D06 and D11. It is recommended to extend the reporting to include experimental results from the remaining seven districts.
• There is an inconsistency in the layout of the document: Figure 5 referred to on line 215 of Page 4, yet it is located on Page 7.
• The caption for Figure 7 is incorrect, and should be corrected to "Edge Dynamics" from "Node Dynamics".
• It is recommended that some recent related studies be discussed in the paper, particularly focusing on their performance with this dataset.

[1] UniST: A Prompt-Empowered Universal Model for Urban ST Prediction. KDD2024.
[2] Fine-Grained Urban Flow Prediction. WWW2021.
[3] When Transfer Learning Meets Cross-City Urban Flow Prediction: Spatio-Temporal Adaptation Matters. IJCAI2022.
[4] Spatio-Temporal Self-Supervised Learning for Traffic Flow Prediction. AAA2023.

1. The technical contribution of this paper is weak. The proposed road network topology generation algorithm is an extension of minimum spanning tree. The primary contribution is the DynST dataset, which may not well align with the expectations for technical novelty at ICLR.
2. The claim made in this paper that "transfer learning tasks in traffic prediction currently lack dedicated datasets and instead rely on datasets designed for non-transfer prediction tasks" seems inaccurate. In fact, there are several multi-city traffic datasets that are widely used for cross-city transfer learning research. This statement overlooks existing efforts in the community that already address this issue.
3. The motivation behind building such a vast traffic dataset spanning 20 years for transfer learning is not clearly justified. For example, the paper does not adequately explain the benefits or rationale for employing such extensive data for transfer learning.
4. From Table 6, it can be seen that expanding the training data from 1 year to 20 years only provides marginal improvements in the model performance in cross-region prediction scenarios. For instance, the MAE (Mean Absolute Error) improves from 26.34 to 26.31 in the zero-shot setting, and from 21.90 to 20.24 in the 3-day setting, which are relatively small gains considering the significant increase in data volume. While the authors emphasize the scale of DynST as a major contribution, the necessity of utilizing such an enormous dataset is questionable.

1. The paper lacks a comprehensive comparison with existing state-of-the-art transfer learning approaches tailored for traffic forecasting.

2. While the authors argue for the benefits of evolving road networks, it is important to consider that the pace of road evolution is relatively slow. This raises questions about the necessity of transferring models over such long time scales. In many cases, retraining models on new road data may be more efficient and yield better performance than relying on transferred knowledge.

(1) Compared to existing works (such as LargeST and STSGCN), this paper's main distinction lies in substantially increasing the volume of collected data. However, since all data is obtained through similar methods from the PEMS system, the contribution is somewhat limited.

(2) All data is sourced exclusively from California's PEMS system, without incorporating data from other regions or countries. More diverse data would be more meaningful for transfer learning tasks.

(3) While the research motivation is to provide rich source datasets for transfer learning in data-scarce target regions, this scenario is not adequately validated in the experimental phase. More realistic scenarios, such as training on California dataset (source dataset D06) and testing transfer learning on Chicago dataset (target dataset), would be more convincing.

(4) Although the paper considers the importance of dynamic road network topology, the dataset construction description and Figure 5 appear to reflect changes in sensors deployment rather than actual road network dynamics. Inferring road network dynamics based on sensors distribution does not align with real-world scenarios.

The proposed tree-based adjacency matrix generation method is introduced but lacks a thorough comparative evaluation against existing methods, particularly the distance-based method.

A deeper exploration of the underlying mechanics of this algorithm, including the rationale behind the choice of distance thresholds and how these thresholds impact connectivity, would provide valuable insight into its effectiveness. Additionally, discussing the computational complexity and efficiency of the proposed method compared to traditional distance-based approaches would help clarify its advantages.

Also, explaining how the adjacency matrix changes to reflect the dynamic nature of the road network—like how often it gets updated and the rules for adding or removing connections—would help readers better understand its real-world importance.