TransferTraj: A Vehicle Trajectory Learning Model for Region and Task Transferability

Thank you for your insightful feedback! We are grateful for your acknowledgment of our work. Below, we address each of your concerns and questions in detail:

[W1] The evaluation of TP task only focuses on the precision of the trajectories’ destinations and ignores the intermediate trajectory, which weakens the claim of accurate trajectory prediction.

We conducted additional experiments evaluating trajectory prediction performance for predicting 1, 2, 3, 4, and 5 future trajectory points on the Chengdu dataset. The experiment results show that the prediction error increases with the number of predicted steps. However, our model consistently outperforms baselines across all prediction lengths, demonstrating its effectiveness in trajectory prediction tasks.

[W2] The unified input and output scheme of three tasks is pivotal to the task transferability, which is just stated in Appendix. Without an explicit, worked example in the main body, it will hurt clarity and reproducibility.

Due to page limitations, the unified input-output scheme details for three tasks are provided in Appendix C. If the paper is accepted, we will move this critical content to the main text for better clarity and reproducibility.

Thank you for your insightful feedback! We are grateful for your acknowledgment of our idea. Below, we address each of your concerns and questions in detail:

[W1] Although several baselines are used for comparison, the paper does not specify how they were trained or tuned. It is not clear whether pre-trained weights were reused or whether hyperparameters were fairly adjusted, which raises concerns about the fairness of the comparisons.

For baseline settings, we adopted hyperparameters reported in the original papers and retrained models from scratch on our datasets. For trajectory representation learning baselines, we first pre-trained their trajectory encoders from scratch, then fine-tuned both encoders and output heads following standard trajectory representation learning protocols. For baselines used in trajectory recovery and origin-destination travel time estimation tasks, we conducted end-to-end training from scratch. Baseline settings are described in Appendix B.2 (lines 448-450), and we will include these clarifications in the revised version.

[W2] The paper lacks implementation details such as the number of layers, embedding size, batch size, and task-specific learning rates.

Due to page limitations, implementation details are provided in Appendix B.2. Our settings are: number of layers $L = 2$, embedding size $d = 128$, batch size 64, and learning rate $1 \times 10^{-4}$. These choices are justified through hyperparameter experiments presented in Appendix G (Figures 3-5). If accepted, we will move this content to the main text.

[Q1] What specific distance metric is used for dis(l, pi)? Is it the Haversine distance, Euclidean, or another geodesic measure appropriate for geographic coordinates?

We use Haversine distance to calculate the distance between the trajectory point and the POI in meters.

[Q2] How to select the hyperparameter of $\varphi_{\text{dist}}^{\text{poi}}$ and $\varphi_{\text{dist}}^{\text{road}}$ ?

The parameters $\varphi_{\text{dist}}^{\text{poi}}$ and $\varphi_{\text{dist}}^{\text{road}}$ are region-specific and adjustable. We added a hyperparameter experiment on Chengdu, Xi'an, and Porto dataset in trajectory prediction task.

The experiment results show that Chengdu and Porto datasets achieve optimal performance at 100 meters, while Xi'an performs best at 150 meters. This difference likely stems from Xi'an's sparser POI and road network distribution (~111.44 POIs and road segments per $\text{km}^2$ in Xi'an vs. ~263.10 in Chengdu and ~282.87 in Porto), where a smaller radius does not capture sufficient spatial context. Performance first improves then declines with increasing distance, indicating that insufficient POIs/road segments provide inadequate contextual information while excessive numbers introduce redundancy. This hyperparameter analysis will be included in the revised version.

[Q3] Regarding the Transformer-based multimodal architecture, it is unclear whether the model was trained entirely from scratch or built upon a pre-existing encoder. Could the authors clarify whether a standard Transformer (e.g., RoFormer, ViT, etc.) was adapted, or whether the encoder was implemented from the ground up? Additionally, how many layers and attention heads are used in the RTTE module?

The model is trained entirely from scratch, adapted from RoFormer by replacing the original rotary position encoding with our learnable spatiotemporal position encoding and substituting the fully connected layer with SC-MoE. These modifications enable better capture of relative relationships between trajectory points and adaptive extraction of movement patterns based on surrounding spatial context, enhancing regional transferability.

According to our hyperparameter experiments (Appendix G, Figure 4), the RTTE model uses 2 layers with 1 attention head.

[Q4] Although Figure 1 provides a schematic overview of the proposed architecture, its connection to the textual description is limited.

We appreciate your careful observation and constructive suggestions. We will revise Figure 1 to better align the visual components with the textual description, providing clearer mapping between the four modalities (spatial, temporal, POI, road network) and their integration in the pipeline. Specifically, we will present a more intuitive pipeline that illustrates the association between trajectory points and the four modalities, as well as the encoding and mixing processes of these modalities. For the TRIE module, we will clearly show its capability to capture the relative relationships between trajectory points. Additionally, we will update the figure’s labels to ensure consistency with the textual description.

Thank you for taking the time to consider the other reviews and rebuttals. We are glad that we were able to answer your questions.

Thank you for your positive and insightful comments. We are delighted that the reviewer found our motivation interesting and reasonable. Below, we address each of your concerns and questions in detail:

[W1 & Q1] The key advantages of TransferTraj compared to traditional trajectory representation learning models.

Trajectory representation learning models pre-train an encoder to obtain universal trajectory embeddings, requiring separate training of task-specific output heads, which incurs additional training costs. Furthermore, our experiments demonstrate that trajectory representation learning requires joint fine-tuning of both the pre-trained encoder and output head when applied to various tasks (as shown in Table 1).

In contrast, TransferTraj eliminates the need for task-specific output heads and additional training, directly adapting to various tasks through our well-designed pre-training scheme while achieving excellent experimental results. Additionally, our model transfers directly to other regions without retraining while maintaining strong performance, whereas existing trajectory representation learning models require separate training for each region.

[W2 & Q2] Further clarify how POI and road network modalities are encoded, and explain why they can be generalized across different regions.

When encoding POIs and road networks, we utilize textual descriptions rather than ID embeddings to capture semantic information. As described in the Preliminaries section, we represent POIs using their name, type, and address as text descriptions, while road segments are described using their name, type, and length. These text descriptions are input into OpenAI's pre-trained text embedding model to generate fixed-dimension vectors. Since the extracted semantic information depends solely on POI and road network attributes and is independent of specific regions, this approach guarantees transferability across different geographical areas.

[Q3] In the task-transferable input-output scheme, how are the masking rates for modalities and trajectory points determined?

During pre-training, we randomly divide each trajectory into multiple non-overlapping segments, with the number of segments set to $0.2 \times T$, where $T$ represents the trajectory length. We then randomly select 40% of these segments and apply trajectory point maskinxg to all points within the selected segments. For the remaining trajectory portions, we further select 20% of the trajectory points and apply either spatial or temporal modality masking. These parameter settings will be clarified in the revised version.

Thank you for your response. Your suggestion is valuable and we will follow your guidance to incorporate the feedback in the revised version.

Thank you for your positive and insightful comments. We are delighted that the reviewer found our motivation interesting and reasonable. Below, we address each of your concerns and questions in detail:

[W1 & Q4] Explaining why TRIE can address the challenge of varying trajectory lengths and the analysis of trajectory length distribution.

About TRIE's generalization to varying trajectory lengths: TRIE achieves length generalization through its learnable spatial-temporal rotation matrix encoding method. Each trajectory point is encoded as a rotation vector on the $d$-dimensional plane, where relative spatial distance information between trajectory points is naturally embedded through cosine and sine functions. This design ensures that angular differences between positions remain both periodic and continuous, maintaining robustness when processing sequences longer than those seen during training. Therefore, TRIE can seamlessly extend to longer sequences during inference while preserving accurate relative position representation and computational stability.

About trajectory length distribution: The datasets exhibit different trajectory length distributions: Chengdu has lengths ranging from 20-120 points with an average distance of 4.057km; Xi'an ranges from 20-120 points with an average distance of 5.253km; Porto ranges from 20-100 points with an average distance of 8.284km. These statistical differences in length distributions make cross-dataset transfer challenging. Our zero-shot transfer experiments demonstrate that TRIE maintains strong generalization ability across these varying distributions, validating its effectiveness.

[W2] The author terms the top-K gating mechanism "prevent the movement patterns from being represented by the same set of experts", however, it does not explain why or how this mechanism successfully enforces expert diversity.

Introducing the top-K gating mechanism in the routing stage can break the model's over-reliance on specific experts, increase the randomness of routing decisions, and improve the overall utilization of experts. In standard MoE architectures, the router tends to repeatedly select a few high-performing experts, causing different movement patterns to be assigned to the same subset of experts. This leads to expert overload and overfitting, weakening the model's expressiveness and generalization ability. The top-K gating mechanism prevents this by forcing the model to explore more diverse expert combinations during training, ensuring better utilization of all available experts. This design enables the model to capture fine-grained differences in movement patterns more effectively. Each expert specializes in a specific subspace of movement patterns, enhancing the model's ability to distinguish and generalize across complex and diverse trajectory behaviors.

[W3] There are some inconsistencies in the writing.

Thank you for your observations. We will revise them in the revision.

[Q1] The experiments are conducted with k=4 and C=8, but Figure 2 shows that the number of activated experts in different-density regions mostly concentrates around 2, 3, and 4. Does this suggest that some experts may be redundant?

Although experts 2, 3, and 4 are frequently activated, each operates within different spatial density contexts. Expert 3 is primarily activated in high-density areas, expert 2 in medium-density areas, and expert 4 in low-density areas. This pattern demonstrates that the model effectively allocates different spatial contexts to different experts through the routing mechanism. Furthermore, the combination of activated experts varies significantly across density levels, confirming that our model adaptively activates appropriate experts based on spatial context, enhancing its modeling capability. Additionally, our hyperparameter study in Figure 5 examines the effect of the number of experts C and activations K on model performance. The experimental results show that reducing the number of experts leads to noticeable performance degradation. Therefore, the current configuration of K=4 and C=8 represents a balanced and optimal choice.

[Q2] When the sampling rate is set to $\mu=16\epsilon$, both the proposed model and the baseline experience a significant increase in error. What might be the underlying cause of this phenomenon?

This phenomenon occurs because the missing rate of trajectory points is extremely high. When $\mu = 16 \times \epsilon$, 93.75% of trajectory points are missing, resulting in severely limited available information. Consequently, both our model and baselines experience degraded performance compared to other sampling rates. However, our model still maintains significant improvement over baselines even under these challenging conditions, further demonstrating its effectiveness.

[Q3] In the pre-training section, how are the lengths of the masked start and end segments defined? how is the modality mask of trajectory points determined?

During pre-training, we randomly divide each trajectory into multiple non-overlapping segments, with the number of segments set to $0.2 \times T$, where $T$ represents the trajectory length. We then randomly select 40% of these segments and apply trajectory point masking to all points within the selected segments. For the remaining trajectory portions, we further select 20% of the trajectory points and apply either spatial or temporal modality masking. These parameter settings will be clarified in the revised version.

Thank you for your response. Your suggestion is valuable and we will follow your guidance to incorporate the feedback in the revised version.