This repository contains the official implementation of our KDD 2026 paper, "MoST: A Foundation Model for Multi-modal Spatio-temporal Traffic Prediction." The MoST model is designed to perform traffic prediction in previously unseen cities using any available extra input modalities.
Experiments were conducted using Python 3.8 and PyTorch 2.4.1 with CUDA 12.4. To install the necessary dependencies, run the following command:
pip install -r requirements.txt
Traffic data were sourced from Urban-Dataset and LargrST. Point-of-interest (POI) data were obtained from the Foursquare. Satellite imagery was acquired from Google Earth.
To train the MoST model, execute the following command. The example configuration file is data_configs/most.yaml. For inference, set the --is_trainingflag to 0.
python run.py --is_training 1 --gpu 0
If you find this work useful for your research, we kindly request citing the following paper:
@inproceedings{xu2026most,
title={MoST: A Foundation Model for Multi-modality Spatio-temporal Traffic Prediction},
author={Xu, Ronghui and Chen, Jihao and Tian, Jingdong and Guo, Chenjuan and Yang, Bin},
booktitle={SIGKDD}
year={2026}
}