Research map / Label-efficient and robust learning / Recognition and long-tail

GTR: A General, Multi-View, and Dynamic Framework for Trajectory Representation Learning

ICML 2025 · Label-efficient and robust learning › Recognition and long-tail · approach: Uncertainty-aware multi-view learning

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 GTR, a pre-train/fine-tune trajectory representation framework that fuses free-space and road-network views, uses a spatio-temporal mixture-of-experts, and updates online by freezing lower Transformer layers.

Problem

Existing trajectory representation learning methods use only a single view (free-space or road network), rely on task-specific training, and do not support online model updates as trajectory data streams in.

Approach

Combines a multi-view encoder (GAT road embeddings, grid+POI embeddings, position and multi-granularity time embeddings), a spatio-temporal MoE that dynamically weights spatial and temporal experts per task, span-masking MLM plus VAE-based hard-negative triplet pre-training, and an online frozen-hot update that freezes the first half of Transformer layers.

Key results

• On Porto travel time estimation, GTR reaches MAE 0.01512 vs. 0.37485 for the best baseline START.
• Top-k similarity search on Beijing: GTR HR@1 0.9906 and MR 1.0130 versus 0.9720 HR@1 for the best baseline JGRM.
• Trajectory imputation on Porto: GTR Recall@3 0.86465 vs. 0.68170 for MTrajRec.

Measured effects reported in the paper

Surprising findings

• Removing the MLM loss causes Porto MR to jump from about 1 to 741, indicating the MLM pre-training is critical for similarity search.
• GTR's training time (179.6 min) is the longest among compared models, exceeding START and JGRM.

Limitations

• Training time is substantially higher than most baselines (179.6 min vs. 5-119 min).
• Evaluated only on two city datasets (Porto and Beijing) in the main experiments, with one extra Chengdu similarity experiment.
• Online updating is evaluated with a single-epoch simulation on the validation set rather than a true streaming deployment.

Experimental setup

Datasets: Porto Beijing

Benchmarks: Porto Beijing

Metrics: MR HR@1 HR@5 MAE MSE MAPE Recall@3 Recall@5 MAP ACC

Baselines: Trembr PIM Toast START LightPath JGRM TS-TrajGen SeqGAN Trajbert Bi-STDDP EB-OTS S3

Builds on: START Trembr Toast JGRM BERT

Closest papers by meaning

• TrajMamba: An Efficient and Semantic-rich Vehicle Trajectory Pre-training Model (NeurIPS 2025 · similarity 0.87)
• PLMTrajRec: A Scalable and Generalizable Trajectory Recovery Method with Pre-trained Language Models (NeurIPS 2025 · similarity 0.87)
• Holistic Semantic Representation for Navigational Trajectory Generation (AAAI 2025 · similarity 0.86)
• Towards Generalizable Trajectory Prediction using Dual-Level Representation Learning and Adaptive Prompting (CVPR 2025 · similarity 0.86)
• UniTR: A Unified Framework for Joint Representation Learning of Trajectories and Road Networks (AAAI 2025 · similarity 0.86)
• GTR-Bench: Evaluating Geo-Temporal Reasoning in Vision-Language Models (ICLR 2026 · similarity 0.86)