Research map / Architectures and efficiency / Pruning and efficient tuning
PLMTrajRec: A Scalable and Generalizable Trajectory Recovery Method with Pre-trained Language Models
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 PLMTrajRec, which fine-tunes a LoRA-adapted BERT-small with natural-language interval prompts, area-flow implicit prompts, and a road-condition passing mechanism to recover map-matched trajectories from sparse GPS data across sampling intervals.
Problem
Map-matched trajectory recovery from sparse GPS trajectories suffers from scarce dense training data, poor generalization across different sampling intervals, and difficulty estimating road conditions at unobserved missing points.
Approach
Uses a pre-trained language model (BERT-small with LoRA) fed with a textual interval/movement prompt, unifies sparse trajectories of different intervals into a common 15-second grid with [m] placeholders, and infers missing-point road conditions by passing area-flow features from neighboring observed points.
Key results
- On Chengdu with 2-minute sampling, PLMTrajRec reaches 81.76% accuracy and 187.4 m MAE versus 78.14% and 197.2 m for the strongest baseline MM-STGED; PLMTrajRec + FT reaches 82.29%.
- On Porto with 1-minute sampling, PLMTrajRec + FT reaches 75.35% accuracy vs 73.16% for MM-STGED.
- In a zero-shot test on unseen 2-minute intervals (trained on 1 and 4 minutes), PLMTrajRec keeps 79.62% accuracy on Chengdu, a 2.14% drop, while baselines drop about 10-18%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add interval-specific fine-tuning | PLMTrajRec · Chengdu and Porto, 4 minutes, 2 minutes, and 1 minute → 15 seconds | accuracy: 1.72% and 0.32% | Helped |
| add the IF-guided explicit trajectory prompt (ablation) | PLMTrajRec · Chengdu, 2 minutes → 15 seconds | accuracy: 81.04 -> 81.76 | Helped |
| add the AF-guided implicit trajectory prompt (ablation) | PLMTrajRec · Chengdu, 2 minutes → 15 seconds | accuracy: 81.43 -> 81.76 | Helped |
| add both trajectory prompts (ablation) | PLMTrajRec · Chengdu, 2 minutes → 15 seconds | accuracy: 80.70 -> 81.76 | Helped |
| add road-network features (ablation) | PLMTrajRec · Chengdu, 2 minutes → 15 seconds | accuracy: 81.28 -> 81.76 | Helped |
| add reference tokens (ablation) | PLMTrajRec · Chengdu, 2 minutes → 15 seconds | accuracy: 80.83 -> 81.76 | Helped |
| use pre-trained BERT instead of randomly initialized BERT (ablation) | PLMTrajRec · Chengdu, 2 minutes → 15 seconds | accuracy: 74.11 -> 81.76 | Helped |
| add the IF-guided explicit trajectory prompt (ablation) | PLMTrajRec · Porto, 2 minutes → 15 seconds | accuracy: 66.13 -> 66.40 | Helped |
| add the AF-guided implicit trajectory prompt (ablation) | PLMTrajRec · Porto, 2 minutes → 15 seconds | accuracy: 66.18 -> 66.40 | Helped |
| add both trajectory prompts (ablation) | PLMTrajRec · Porto, 2 minutes → 15 seconds | accuracy: 65.52 -> 66.40 | Helped |
| add road-network features (ablation) | PLMTrajRec · Porto, 2 minutes → 15 seconds | accuracy: 66.11 -> 66.40 | Helped |
| add reference tokens (ablation) | PLMTrajRec · Porto, 2 minutes → 15 seconds | accuracy: 65.89 -> 66.40 | Helped |
Surprising findings
- Randomly initialized BERT drops Chengdu accuracy from 81.76% to 74.11%, so the pre-trained weights matter substantially but the gap to baselines is partly due to the architecture.
- Reported Porto gains over baselines are small (e.g., 66.40% vs 65.69% accuracy at 2 minutes), and the Porto Acc/Recall of PLMTrajRec are not best on some metrics.
Limitations
- Model has the largest parameter count (176 MB) among compared methods, though training and inference are faster.
- Evaluation limited to two taxi datasets from Chengdu and Porto with a fixed 15 s target interval.
- Road-condition features rely on traffic flow counts aggregated on a 64x64 grid with hourly slices, whose sensitivity is not deeply analyzed.
Experimental setup
Datasets: Chengdu (DiDi) Porto (Kaggle taxi) OpenStreetMap road network
Benchmarks: Chengdu Porto
Metrics: Accuracy Recall Precision MAE RMSE
Baselines: HMM + ShortestPath Linear + HMM MPR + HMM DHTR + HMM AttnMove + Rule MTrajRec RNTrajRec MM-STGED
Builds on: BERT LoRA MTrajRec Learnable Fourier Features
Closest papers by meaning
- GTR: A General, Multi-View, and Dynamic Framework for Trajectory Representation Learning (ICML 2025 · similarity 0.87)
- TrajMamba: An Efficient and Semantic-rich Vehicle Trajectory Pre-training Model (NeurIPS 2025 · similarity 0.86)
- Pedestrian Trajectory Prediction with Missing Data: Datasets, Imputation, and Benchmarking (NeurIPS 2024 · similarity 0.86)
- Generative Human Trajectory Recovery via Embedding-Space Conditional Diffusion (ICML 2025 · similarity 0.85)
- ProDiff: Prototype-Guided Diffusion for Minimal Information Trajectory Imputation (ICML 2025 · similarity 0.85)
- Bridging Traffic State and Trajectory for Dynamic Road Network and Trajectory Representation Learning (AAAI 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.