A discrete trajectory autoencoder based on language modeling

This repository contains the training scripts for the discrete trajectory autoencoder developed at Cardiff.

framework

In this repository:

• Inspired by language modeling for discrete data, we treat urban mobility trajectory data as a form of language (discrete sequences).
• We first pretrain a basic discrete, road-segment-level trajectory autoencoder based on a BART structure.
• After pretraining, we perform autoencoder compression to obtain denser embeddings of trajectory data.

git clone https://github.com/urban-mobility-generation/Discrete-Trajectory-Autoencoder.git
cd Discrete-Trajectory-Autoencoder

Make sure you have prepared the trajectory dataset.
We will provide a preprocessed public dataset from Porto soon!

📁 Dataset Overview Processed Porto Dataset

The folder contains:

• all_traj_results.npy: Normalized trajectory data
• all_attr_results.npy: Normalized attributes for each trajectory (e.g., departure time, distance, origin/destination grid)
• edge_map_feature_po.csv: Road network information
• final_segments_all_train_data.pkl: Segment-level trajectories

Please follow the usage requirements set by the original owners of this open-source dataset!

nohup python pretrain_mb.py > train_mask_e4d2_mask.log 2>&1 &

nohup python train_latent_model_coor.py > trainl16d128_mask_coor_04.log 2>&1 &

The trained autoencoder can be used in Cardiff for mobility generation and also has broad applications such as trajectory representation, imputation, and infilling tasks.

@article{guo2025leveraging,
 title={Leveraging the Spatial Hierarchy: Coarse-to-fine Trajectory Generation via Cascaded Hybrid Diffusion},
 author={Guo, Baoshen and Hong, Zhiqing and Li, Junyi and Wang, Shenhao and Zhao, Jinhua},
 journal={arXiv preprint arXiv:2507.13366},
 year={2025}
}

The training pipeline of the compression module is inspired by the following projects:

• Latent Diffusion for Language Generation
• Flamingo

We thank the authors for their high-quality implementations and open-source contributions.