STMetaT

Spatio-Temporal Meta-Learning for Trajectory Representation Learning

Knowledge-Based Systems, Volume 327, Article 114141 (2025)

Zhouzheng Xu · Yuxing Wu · Hang Zhou · Chaofan Fan · Bingyi Li · Kaiyue Liu · Yaqin Ye · Shunping Zhou · Shengwen Li

---

TL;DR: STMetaT learns trajectory representations across heterogeneous spatial and temporal contexts. It constructs constrained meta-learning tasks, fuses GPS, road, and POI views, and transfers the resulting representation to destination prediction, travel-time estimation, and similar-trajectory search.

Highlights

• Spatio-temporal task construction. Support/query subsets are sampled under spatial-grid and temporal-window constraints.
• Multi-view trajectory encoding. GPS coordinates, road semantics, and POI context are fused instead of relying on a single trajectory view.
• Meta-learning for heterogeneity. Inner-loop task adaptation and outer-loop optimization target transfer across distinct spatial and temporal conditions.
• Multi-task evaluation. The pipeline contains prediction and retrieval heads for destination prediction (DP), travel-time estimation (TTE), and similar-trajectory search.

Method at a Glance

Stage	Purpose	Code
1. Constrained task sampling	organize trajectories by spatial grid and temporal window	meta_maml_train.py, data.py
2. Multi-view encoding	combine geometric, temporal, road, and POI signals	models/encode.py, models/traj_clip.py
3. Sequence modeling	model long-range trajectory dependencies	models/mamba/, models/mamba2/
4. Adaptation and evaluation	fine-tune task heads and evaluate prediction or retrieval	pipeline.py, models/predictor.py

Installation

A CUDA-enabled environment is recommended for the Mamba kernels. The code imports the following core packages:

pip install torch numpy pandas tables scikit-learn einops tqdm higher packaging
pip install mamba-ssm causal-conv1d

Data and Configuration

Experiments are driven by settings/<name>.json. Each setting identifies:

train_traj_df   HDF5 trajectory table (key: trips)
test_traj_df    HDF5 trajectory table (key: trips)
poi_df          HDF5 POI table (key: pois)
poi_embed       NumPy POI embedding matrix
road_embed      NumPy road embedding matrix

settings/local_test.json points to the small files under samples/. For full experiments, store the datasets outside Git and update the paths in a dedicated settings file.

Optional output locations are controlled through environment variables:

export SETTINGS_CACHE_DIR=/path/to/settings-cache
export MODEL_CACHE_DIR=/path/to/checkpoints
export PRED_SAVE_DIR=/path/to/predictions
export SEARCH_META_DIR=/path/to/search-metadata

Quick Start

Run the configured pretraining, fine-tuning, and evaluation pipeline:

python main.py --settings local_test --cuda 0

The active stages come directly from the selected JSON file:

• pretrain: self-supervised trajectory representation learning;
• finetune: downstream adaptation with the configured task head;
• test: dp, tte, or search evaluation and optional prediction export.

Entry Points

Entry point	Role
main.py	configuration-driven pretrain → fine-tune → test pipeline
meta_maml_train.py	constrained support/query task construction and MAML-style training

The sample quick start runs main.py. Formal STMetaT meta-learning experiments use the dedicated meta-learning entry point and a settings file containing meta_lr, inner_lr, num_inner_steps, k_shot, and q_shot.

Outputs

Path	Content
$SETTINGS_CACHE_DIR/<timestamp>.json	immutable copy of the selected setting
$MODEL_CACHE_DIR/*.pretrain	pretrained trajectory encoder
$MODEL_CACHE_DIR/*_trajclip.finetune	fine-tuned trajectory encoder
$MODEL_CACHE_DIR/*_predhead.finetune	fine-tuned downstream head
$PRED_SAVE_DIR/<run>/	optional predictions and targets
$SEARCH_META_DIR/<dataset>/	generated retrieval candidates and metadata

Citation

If STMetaT is useful in your research, please cite:

@article{xu2025stmetat,
  author  = {Xu, Zhouzheng and Wu, Yuxing and Zhou, Hang and Fan, Chaofan and Li, Bingyi and Liu, Kaiyue and Ye, Yaqin and Zhou, Shunping and Li, Shengwen},
  title   = {Spatio-Temporal Meta-Learning for Trajectory Representation Learning},
  journal = {Knowledge-Based Systems},
  volume  = {327},
  pages   = {114141},
  year    = {2025},
  doi     = {10.1016/j.knosys.2025.114141}
}