acceptodds
Under review as a conference paper at ICLR 2027

TEMPT: Trajectory Mixing Decomposition for Low-resource GPS Trajectory Prediction

Abstract

Trajectory prediction plays a pivotal role in diverse applications, including intelligent transportation, urban planning, and autonomous navigation. However, prevailing trajectory prediction methods, particularly Transformer-based approaches demand substantial computational resources, making them unsuitable for resource-constrained environments like edge devices, mobile platforms, and embedded systems. To address this challenge, we propose TEMPT (Trajectory Mixing Decomposition), a lightweight yet highly accurate framework for GPS trajectory prediction. TEMPT employs a decomposition-inspired stacked architecture with lightweight MLP-Mixer blocks to extract multi-scale trajectory components, capturing both coarse and fine-grained patterns. Additionally, a dynamic weighted aggregation mechanism fuses adaptive patch representations into a unified embedding space, further enabling efficient handling of variable-length inputs. Extensive experiments on four trajectory datasets show that TEMPT achieves state-of-the-art accuracy while reducing computational costs by 56 compared to conventional approaches. It further sustains real-time single-trajectory inference on one CPU thread, so a dedicated accelerator is not required at deployment time. The source code and trained models will be made publicly available.

Primary area:
learning on time series and dynamical systems
Keywords:
GPS Trajectory, Trajectory Prediction, Low-resource Computing

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_Rs6KOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces an efficient trajectory prediction method for low-resource GPS trajectory prediction. It is designed to capture large-scale, medium-scale, and fine-scale motion patterns through different layers, and the final prediction is formulated as the sum of predictions from all layers. The idea is interesting, but the authors simplify the problem by predicting solely based on an agent's past geographical coordinates. This is a potentially limiting simplification, as road networks, agent interactions, and other contextual information are important for trajectory prediction. Without considering them, prediction accuracy could be compromised. Therefore, the efficiency may come with a trade-off in prediction accuracy in more complex real-world scenarios.

Critical strengths
  • The proposed method is efficient, supports deployment on mobile devices, and can handle varying lengths of historical trajectories as input.
  • The authors conducted experiments on multiple datasets, including real-world trajectory datasets collected in different cities and a simulated UAV dataset. The comparisons also cover different prediction horizons.
  • The authors conducted ablation studies to evaluate the contributions of key components, including multi-scale decomposition, dynamic weighted aggregation, and residual modeling. The results demonstrate the effectiveness of these components in improving prediction accuracy.
Critical weaknesses
  • The authors simplify the trajectory prediction problem by relying solely on the target agent's past latitude and longitude coordinates. However, a significant amount of previous work has focused on the importance of considering road networks, neighboring agents, and environmental context. This simplification may reduce prediction accuracy and therefore limit the method's feasibility in practical applications. It may also raise fairness concerns in the experimental comparisons, as some of the baseline methods were originally designed to model richer structural information. The authors' adaptations may no longer preserve the original strengths of these methods.

  • Although the authors claim that map information, neighboring agents' interactions, traffic signals, etc., can be integrated into the method, they do not explain in detail how this would be achieved or conduct experiments to verify its effectiveness. Such information is important for real-world trajectory prediction applications and should ideally be considered in the current study rather than left entirely as future work.

  • From Table 1, it appears that using a short trajectory (the past 30 observations) as input generally achieves better prediction accuracy than using a long trajectory (the past 400 observations), even for the same prediction horizon. The variable-length input setting also does not consistently outperform the short-input setting. However, these settings may not not directly comparable because they may involve different input configurations and evaluation samples. This raises the question of what benefits the ability to process varying lengths of input trajectories provides beyond flexibility, and whether longer historical trajectories actually improve prediction accuracy.

  • The proposed method directly uses absolute longitude and latitude coordinates as input and predicts future absolute positions. However, existing trajectory prediction methods have demonstrated the benefits of modeling relative displacements or velocities, which can better capture motion dynamics and reduce dependence on absolute geographical locations. The authors do not investigate whether alternative trajectory representations, such as relative coordinates, velocity, or displacement-based prediction, could further improve performance. Such representations could be used as both inputs and prediction targets.

  • The authors claim that different layers capture motion patterns at different temporal scales, from global trends to fine-grained variations. However, it is unclear whether the proposed decomposition actually achieves this. While the ablation studies show that the decomposition improves prediction accuracy, they do not explain what each layer has learned. It would be helpful to visualize the decomposed trajectory components and analyze whether they indeed capture different temporal patterns.

Questions for authors
  • How can the proposed method incorporate additional information, such as interactions between neighboring agents and road network information from a digital map? Can the proposed method still outperform existing trajectory prediction methods after integrating such contextual information?

  • What is the advantage of enabling the model to handle varying lengths of historical trajectories? Does using a shorter input trajectory consistently lead to better prediction accuracy? Could the authors further analyze the impact of input trajectory length under the same evaluation setting?

Reviewer_LpokOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes TEMPT, a lightweight framework for GPS trajectory prediction in resource-constrained environments. The method employs a hierarchical residual decomposition architecture with MLP-Mixer blocks to extract multi-scale trajectory representations. Experiments on four trajectory datasets demonstrate competitive prediction accuracy with relatively low computational costs, highlighting the potential of TEMPT for resource-constrained deployment.

Critical strengths

(1) The proposed framework combines multi-scale patch processing with dynamic weighted aggregation, allowing the model to handle trajectories of varying lengths without modifying its parameter dimensions.

(2) The paper evaluates TEMPT across multiple datasets and prediction horizons using several accuracy metrics.

(3) TEMPT achieves strong prediction accuracy with a relatively small model size and low computational complexity.

Critical weaknesses

(1) The claim that coordinate feature mixing preserves heading and speed lacks direct experimental support. Improvements in prediction accuracy alone do not demonstrate that the model explicitly captures or preserves kinematic quantities.

(2) The residual whiteness regularization relies on assumptions about the statistical properties of the residual signal, particularly its temporal decorrelation. However, GPS trajectories are inherently non-stationary motion signals, involving acceleration, deceleration, turning, stopping, and starting.

(3) The experimental results indicate that residual decomposition and regularization improve prediction accuracy and reduce correlations among the extracted components. However, these findings only demonstrate the effectiveness of the proposed structure as a prediction regularization strategy.

(4) The authors are encouraged to include comparisons with state-of-the-art methods specifically designed for variable-length trajectory prediction.

Questions for authors

See Weaknesses.

Reviewer_HvAbOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper proposes a lightweight model for single-agent GPS trajectory forecasting from the agent's own coordinate stream, short as TEMPT. TEMPT stacks L layers in a doubly residual scheme. Each layer patches the current residual at its own scale and mixes features within each patch. It subtracts a reconstructed history component from the residual. It also pools the patches with softmax weights and maps the result to a future component. The forecast is the sum of these components. A residual loss from MSD-Mixer regularizes the final residual. No weight depends on the number of patches, so one model accepts histories of any length, which is actually important if deploy this system to the real production env.

Experiments use two DiDi taxi datasets, Chengdu and Xi'an, Porto data, and a simulated UAV dataset. These are public wide-use benchmarks. The exps cover short-history, long-history, and variable-length scenarios. The paper reports the lowest RMSE and MAE among baselines of main exp, with fewer parameters. Also, important to really deploy in production.

Critical strengths

S1. A practical problem with a clear target. The paper targets on-device forecasting from a vehicle's or drone's own GPS stream. On such devices, compute is limited and history lengths vary. Heavy models and fixed-length models both fit this setting poorly. TEMPT decomposes the history layer by layer, from coarse to fine patch scales, and sums the per-layer forecasts. It replaces cross-patch mixing with a softmax pooling over patches. No weight depends on the number of patches, so one model serves any history length without padding or truncation. The design is simple, and the shape table and pseudocode make this property easy to check.

S2. Strong results with thorough reporting. TEMPT gives the best accuracy on taxi data from several cities and on simulated UAV data. This holds under short, long and variable-length histories. It also uses fewer parameters and less computation than the baselines. The ablations remove each component in turn, and the comparison of aggregation rules includes a useful permutation control. The appendix is unusually thorough. It adds displacement metrics, analytical and decomposition baselines, and held-out driver, period and city splits. It also covers cross-domain transfer, uncertainty heads, incomplete histories, hardware latency, and seed and bootstrap statistics. The protocol is clearly documented, including trip-level splits, a shared tuning budget and the latency measurement rule.

Critical weaknesses

W1. Use more suitable baselines and state the novelty more precisely. Several main baselines come from grid or sensor traffic forecasting, or from trajectory generation. They are adapted by treating longitude and latitude as two graph nodes, which removes the role of their spatial modules. Some results suggest under-tuning. A linear baseline trails simply holding the last position, and simple sequence models beat several main baselines. Native GPS trajectory models on the same cities, such as TransferTraj and UniTraj, would make the comparison more convincing. A short description of how each baseline was adapted would also help. Relatedly, the contribution relative to prior work could be stated more clearly. Single-vehicle GPS forecasting on these cities has been studied by TransferTraj, UniTraj and UVTM. Short or variable observation lengths have been addressed for pedestrian trajectories by FlexiLength and by Monti et al. The layer recursion, summed per-layer forecasts, descending patch sizes and residual loss closely follow MSD-Mixer, which in turn builds on N-BEATS. The new pooling step is attention-based MIL pooling (Ilse et al.) with a position feature added. A precise statement of what is new, beyond this combination, would help readers judge the contribution. A discussion of N-HiTS, TimeMixer and variable-context patch models such as TimesFM and Moirai would also place the work better. This is the main concern for the paper.

W2: Strengthen the efficiency measurements. The latency figures may be dominated by overhead unrelated to the model. CPU latency grows with the horizon, although only the linear head depends on it. A single-layer linear baseline runs slower than TEMPT, and CPU and GPU timings are nearly identical. A per-operator profile and batched throughput would help. So would a clear statement of what the timed region includes.

Questions for authors
  1. Units of the reported errors. The error magnitudes seem larger than the standardized space of Appendix A.2 would allow. Could the authors report results in meters? This would make the gains easier to judge.
  2. Parameter and FLOP counts. I could not reproduce the reported counts from the stated widths. Could the authors give a per-stage breakdown?
  3. Latency measurement. CPU latency grows with the horizon, and a single linear baseline runs slower than TEMPT. Could the authors clarify what the timing includes, or add batched throughput?
  4. Baselines and positioning. Several baselines were designed for traffic grids, sensors or trajectory generation. Could the authors add native GPS models such as TransferTraj or UniTraj, and state what is new relative to MSD-Mixer?
  5. Patch alignment and order sensitivity. Patches start at the oldest point, and the model is quite sensitive to patch order. Could the authors explain this choice, or try aligning patches to the newest point?
  6. Evaluation details. Could the authors clarify the statistical tests in Appendix B.10 and B.11, the number of test windows per scenario, and whether ablations use validation or test data?

Overall, I find the problem relevant and the experiments extensive. I would be happy to raise my score if these points are addressed in the rebuttal.

2 comments

Dear Reviewer,

I hope this message finds you well. As the discussion period is nearing its end, I wanted to ensure we have addressed all your concerns satisfactorily.

Please let us know if any point needs further clarification or if you have additional questions. Your insights are invaluable to us, and we will respond promptly.

Thank you again for your time and thoughtful feedback.

Thank you very much for the detailed response and the additional experiments. I appreciate the effort the authors have put into the rebuttal. Several points are now clearer to me, including the patch alignment, the bootstrap unit, the evaluation splits, and the positioning relative to MSD-Mixer. I also appreciate the planned discussion of the additional related work.

After careful consideration, I would like to keep my original score for now. A few concerns remain open. The parameter breakdown does not fully match the architecture described in the paper. The efficiency comparison is hard to interpret without forward-only latency for the baselines. The new TransferTraj results also do not match Table 1 for the same setting. I believe the problem is relevant and the experiments are extensive, and I hope these suggestions help improve the paper.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.