Research map / Label-efficient and robust learning / Recognition and long-tail
UniTR: A Unified Framework for Joint Representation Learning of Trajectories and Road Networks
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 UniTR, a contrastive framework that jointly learns road network segment and trajectory representations via hierarchical propagation between a road encoder and trajectory encoder, trained with a triple-level contrastive loss and proximity/mobility-aware sampling.
Problem
Joint representation learning of road networks and trajectories, where existing methods model them in separate branches and fail to capture many-to-many interactions and cross-scale embedding optimization.
Approach
A hierarchical propagation mechanism (road-to-trajectory aggregation and trajectory-to-road feedback) models many-to-many interactions, combined with a triple-level (segment, trajectory, cross-level) contrastive loss using proximity- and mobility-aware positive/negative sampling.
Key results
- On road type classification in Porto, UniTR reaches Mi-F1 0.638 and Ma-F1 0.564, above the best baseline JCLRNT (0.593/0.402).
- On similar trajectory search in Porto, UniTR achieves MR 1.740 and HR@10 0.984 versus JCLRNT's 5.003 and 0.939.
- Travel time estimation in Porto improves to MAE 144.6 versus 159.3 for JCLRNT.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add visual features from street-view images (ablation) | UniTR · Road Type Classification, Porto | Mi-F1: 0.613 -> 0.638 | Helped |
| add visual features from street-view images (ablation) | UniTR · Traffic Speed Inference, Porto | MAE: 1.978 -> 2.101 | Mixed |
| add trajectory-to-road propagation (ablation) | UniTR · Traffic Speed Inference, Porto | MAE: 2.144 -> 2.101 | Helped |
| add trajectory-level contrastive loss (ablation) | UniTR · Travel Time Estimation, Porto | MAE: 146.6 -> 144.6 | Helped |
| add cross-level contrastive loss (ablation) | UniTR · Road Type Classification, Porto | Mi-F1: 0.633 -> 0.638 | Helped |
| add cross-level contrastive loss (ablation) | UniTR · Travel Time Estimation, Porto | MAE: 142.4 -> 144.6 | Mixed |
Surprising findings
- Removing visual features (w/o vis) slightly improves travel time MAE in Porto (1.978 vs 2.101 for traffic speed inference), showing ablations are not uniformly worse.
Limitations
- Evaluated only on two cities (Porto and Chengdu), limiting generality.
- Relies on map matching of GPS trajectories to road segments and on street view images being available for each segment.
- Model complexity grows with road network scale and trajectory batch size.
Experimental setup
Datasets: Porto (Kaggle 2015) Chengdu (Didi Chuxing, private) OpenStreetMap road networks Google Maps and Baidu Maps street view images
Benchmarks: Porto Chengdu
Metrics: Mi-F1 Ma-F1 MAE RMSE MR HR@10
Baselines: Node2Vec SRN2Vec RFN HRNR SARN JCLRNT
Builds on: Chebyshev graph convolution JCLRNT Toast SwinTransformer
Closest papers by meaning
- Road Network Representation Learning with the Third Law of Geography (NeurIPS 2024 · similarity 0.88)
- Bridging Traffic State and Trajectory for Dynamic Road Network and Trajectory Representation Learning (AAAI 2025 · similarity 0.87)
- GTR: A General, Multi-View, and Dynamic Framework for Trajectory Representation Learning (ICML 2025 · similarity 0.86)
- Urban Region Embedding via Multi-View Contrastive Prediction (AAAI 2024 · similarity 0.84)
- Towards Dynamic Spatial-Temporal Graph Learning: A Decoupled Perspective (AAAI 2024 · similarity 0.84)
- Holistic Semantic Representation for Navigational Trajectory Generation (AAAI 2025 · similarity 0.84)
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.