Research map / Label-efficient and robust learning / Zero-/few-shot and meta-learning

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

NeurIPS 2025 · Label-efficient and robust learning › Zero-/few-shot and meta-learning · approach: Zero-shot robustness and transfer · 1 citations (OpenAlex)

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 TransferTraj, a pre-trained vehicle GPS trajectory model with a region-transferable encoder (RTTE with TRIE and SC-MoE) and a masking/recovery input-output scheme enabling zero-shot task transfer and cross-region transfer.

Problem

Trajectory learning models are typically trained per region and per task, and existing transferable embedding methods transfer poorly across regions and need retrained prediction heads for each task.

Approach

A trajectory encoder combining spatial, temporal, POI and road-network modalities, with rotary-style relative spatial attention (TRIE) and a spatial-context mixture-of-experts (SC-MoE), plus a unified masking-and-recovery input-output scheme so one pre-trained model handles trajectory prediction, recovery and OD travel time estimation without task-specific heads.

Key results

• Pre-training alone (wo ft) gives average 20.18%, 17.87%, 7.94% gains over SOTA baselines START, MM-STGED, DOT on TP, TR, OD TTE respectively.
• Zero-shot region transfer on trajectory prediction improves by an average 83.70% over SOTA; few-shot TP improves 33.68%.
• Ablation: removing SC-MoE causes a 14.52% drop on OD TTE in the Chengdu dataset.

Measured effects reported in the paper

Limitations

• Cross-region classification tasks such as trajectory-user linking or destination road segment prediction are not handled due to varying users and road networks across regions.
• Baselines for trajectory recovery are adapted to coordinate outputs, and zero-shot recovery transfer was not evaluated because road-ID classification does not transfer.

Experimental setup

Datasets: Chengdu (DiDi) Xi'an (DiDi) Porto (ECML/PKDD 2015) AMap POI data OpenStreetMap road networks

Benchmarks: Chengdu Xi'an Porto

Metrics: RMSE MAE MAPE

Baselines: t2vec Trembr CTLE Toast TrajCL START LightPath DHTR MTrajRec RNTrajRec MM-STGED DOT

Builds on: RoFormer noisy top-K gating mixture of experts START Trembr

Closest papers by meaning

• TrajMamba: An Efficient and Semantic-rich Vehicle Trajectory Pre-training Model (NeurIPS 2025 · similarity 0.88)
• GTR: A General, Multi-View, and Dynamic Framework for Trajectory Representation Learning (ICML 2025 · similarity 0.85)
• UniTraj: Learning a Universal Trajectory Foundation Model from Billion-Scale Worldwide Traces (NeurIPS 2025 · similarity 0.85)
• STraj: Self-training for Bridging the Cross-Geography Gap in Trajectory Prediction (AAAI 2025 · similarity 0.85)
• PLMTrajRec: A Scalable and Generalizable Trajectory Recovery Method with Pre-trained Language Models (NeurIPS 2025 · similarity 0.84)
• Towards Generalizable Trajectory Prediction using Dual-Level Representation Learning and Adaptive Prompting (CVPR 2025 · similarity 0.84)