A Trajectory Probability Network for City-Scale Road Volume Prediction

1. The framing of the problem—traffic volume prediction—appears very similar to traffic flow prediction, which is a prominent topic in spatiotemporal data mining and intelligent transportation systems. Given this context, it seems unusual that the proposed method is not compared with any of the numerous existing traffic flow prediction methods.
2. The presentation of challenges could be improved. The paper states that "the majority of these methods provide only deterministic reconstructions of missing data, overlooking inherent uncertainty in other potential scenarios." The term "inherent uncertainty" is somewhat vague and lacks further elaboration. Additionally, since MAE is used as the main evaluation metric in experiments, the probabilistic prediction capability of the proposed method isn't thoroughly assessed. This raises questions about whether the method can truly account for uncertainty in traffic comprehensively.
3. Another challenge mentioned is that "in some underdeveloped areas, observational data can be extremely sparse." However, it remains unclear how this challenge is addressed by the proposed method.
4. It seems that predicting future traffic volumes requires first simulating each vehicle's appearance probability in the city, raising concerns about computational efficiency—an aspect not assessed in experiments.
5. Some implementation details are missing. For instance, the calculation of adjacency tables $A$ is not clearly defined, leaving ambiguity regarding how "the other adjacency tables" are calculated and aligned with $A_0$.

W1: The paper is hard to follow due to an unclear problem definition. It is not clear whether the focus is on predicting future traffic volumes or on trajectory recovery. In Section 3.2, it is also unclear whether T for Y represents past or future time steps. Additionally, the baselines address trajectory recovery, while the related works in Section 2.1 focus on future prediction.

W2: The descriptions of deep learning methods and checkpoint-based methods are confusing. Checkpoint-based methods appear to be a subset of deep learning methods, with the main difference from existing works lying in the data collection approach. However, it is unclear why checkpoint-based methods are considered “more applicable to real-world scenarios.”

W3: The trajectory representation may involve a high level of repetition. For example, repeating node 2 four times to indicate that the car spends four time steps on (2, 3) could result in substantial computational overhead and memory usage.

W4: Only MAE is reported as a performance metric. Additional metrics like RMSE and MAPE would provide a more comprehensive evaluation.

W5: The transition from Equation 6 to Equation 7 in the loss function is not clearly explained, especially regarding how the simplification occurs.

W6: The ablation study shows that using multi-head attention (8) yields an MAE of 0.538, significantly outperforming multi-query attention with an MAE of 0.667. Similarly, configuration (6) with the BVLC embedding shape achieves an MAE of 0.451, notably better than the default setting (0.667), contradicting the paper’s claim that the default setting “has little effect on performance.”

1. There is notable redundancy in the paper, particularly in the related work section (e.g., the first two paragraphs and Section 2.1). Reducing this duplication could improve readability.
2. The descriptions of trajectory interpolation/generation and traffic volume prediction are somewhat unclear. When the authors talk about predicting in a single step, they seem to be talking about autoregressive trajectory generation, while the task they focus on is traffic volume prediction.
3. Following 2, the relevance of the baselines, Cam-Traj-Rec and Traj2Traj, to the traffic volume prediction task is unclear, as these primarily address trajectory interpolation/generation. In addition, when considering trajectory generation, more advanced baselines need to be compared, e.g., MobilityGPT[1], and TS-TrajGen[2], which also show similar trajectory maps like Figure 5 in the paper.
4. The experimental dataset is insufficient. The data on Boston are all generated shortest path data, which may not fully capture real-world conditions. Given TraPNet’s use of cross-attention in the node dimension, I am concerned about the performance and efficiency of the model on larger road networks (10k, or 100k+ nodes).

[1] Haydari, Ammar, et al. "Mobilitygpt: Enhanced human mobility modeling with a gpt model." [2] Jiang, Wenjun, et al. "Continuous trajectory generation based on two-stage GAN." Proceedings of the AAAI Conference on Artificial Intelligence.

1. Missing intuitions and experiments. I think more experiments and validation should be added to explain why the current design outperform SOTA baselines by a large margin. It is difficult to understand based on current explanations and experiments, which only have some numbers, but not concrete evidence of why.
2. More most recent baselines should be added to validate the effectiveness.