Research map / Label-efficient and robust learning / Recognition and long-tail
TrajMamba: An Efficient and Semantic-rich Vehicle Trajectory Pre-training Model
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes TrajMamba, a Mamba2-based vehicle trajectory pre-training model that jointly models GPS and road views, injects travel purposes via contrastive alignment with road/POI text views, and learns a mask generator for compressing redundant points through knowledge distillation.
Problem
Learning travel semantics (movement patterns and travel purposes) from vehicle GPS trajectories is costly when language-model POI/road text is integrated, and high-frequency trajectories contain redundant points that hurt efficiency and embedding quality.
Approach
A Mamba2-style encoder with GPS-SSM and Road-SSM branches whose selection parameters are driven by high-order movement features; travel purpose is distilled into the encoder via InfoNCE alignment with road and POI text views (textual encoders used only in pre-training); a learnable soft mask generator selects key points and a student encoder is trained with MEC loss to match the teacher's full-tra
Key results
- On Chengdu destination prediction, TrajMamba reaches 129.47 m GPS RMSE and 58.21% road Acc@1 versus 215.99 m and 54.27% for the second-best JGRM.
- Arrival time estimation on Chengdu under fine-tuning yields RMSE 50.17 s vs 65.24 s for JGRM.
- Similar trajectory search on Chengdu reaches 96.53% Acc@1 vs 94.45% for TrajCL, with embedding time 1.721 s vs 94.048 s for JGRM.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace Traj-Mamba blocks with vanilla Mamba2 blocks (ablation) | TrajMamba · Chengdu DP GPS | RMSE (meters): 129.47 -> 152.51 | Hurt |
| replace Mamba-based modules with vanilla Transformer layers (ablation) | TrajMamba · Chengdu DP GPS | RMSE (meters): 129.47 -> 182.00 | Hurt |
| add travel purpose-aware pre-training instead of reconstruction pre-training (ablation) | TrajMamba · Chengdu DP GPS | RMSE (meters): 153.84 -> 129.47 | Helped |
| add knowledge distillation pre-training (ablation) | TrajMamba · Chengdu DP GPS | RMSE (meters): 145.34 -> 129.47 | Helped |
| add learnable trajectory compression with the mask generator (ablation) | TrajMamba · Chengdu DP GPS | RMSE (meters): 174.65 -> 129.47 | Helped |
| replace Douglas-Peucker compression with the learnable mask generator (ablation) | TrajMamba · Chengdu DP GPS | RMSE (meters): 173.95 -> 129.47 | Helped |
| replace direct downsampling at a 60% sampling ratio with the learnable mask generator (ablation) | TrajMamba · Chengdu DP GPS | RMSE (meters): 150.70 -> 129.47 | Helped |
| add rule-based filtering preprocessing (ablation) | TrajMamba · Chengdu DP GPS | RMSE (meters): 125.56 -> 129.47 | Hurt |
Surprising findings
- Removing the mask generator's rule-based filtering (w/o Filter) performs comparably to or slightly better than the full model on several metrics, suggesting the learnable soft mask does the work.
- TrajMamba does not have a particularly short training time compared with baselines such as MMTEC, which trains faster (0.378 min/epoch).
Limitations
- Relies on road segment and POI semantics that differ across cities, so cross-city transfer and zero-shot evaluation are not feasible.
- Training time is not especially short compared with some baselines.
- Evaluated only on two Didi taxi datasets from China (Chengdu and Xian).
Experimental setup
Datasets: Chengdu (Didi) Xian (Didi) AMap POIs OpenStreetMap road networks
Benchmarks: Chengdu Xian
Metrics: RMSE MAE Acc@1 Acc@5 Recall MAPE Mean Rank model size train time embed time
Baselines: t2vec Trembr CTLE Toast TrajCL LightPath START MMTEC JGRM
Builds on: Mamba2 InfoNCE MEC loss JGRM START
Closest papers by meaning
- Holistic Semantic Representation for Navigational Trajectory Generation (AAAI 2025 · similarity 0.89)
- TransferTraj: A Vehicle Trajectory Learning Model for Region and Task Transferability (NeurIPS 2025 · similarity 0.88)
- GTR: A General, Multi-View, and Dynamic Framework for Trajectory Representation Learning (ICML 2025 · similarity 0.87)
- PLMTrajRec: A Scalable and Generalizable Trajectory Recovery Method with Pre-trained Language Models (NeurIPS 2025 · similarity 0.86)
- Trajectory Mamba: Efficient Attention-Mamba Forecasting Model Based on Selective SSM (CVPR 2025 · similarity 0.86)
- TrajAgent: An LLM-Agent Framework for Trajectory Modeling via Large-and-Small Model Collaboration (NeurIPS 2025 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.