ICML 2025: Robust Spatio-Temporal Centralized Interaction for OOD Learning

This is the official repository of our ICML 2025 paper. Please note that our code repository only provides some of the pre-processed, directly runnable data for KnowAir and TrafficStream. The raw data for LargeST needs to be obtained and processed as follows, with the corresponding processing files attached in the repository. If not available, they can be obtained from the official repository of the respective dataset.

1. Introduction about the Datasets

1.1 Generating SD, GBA, GLA, CA datasets from LargeST dataset for OOD Setting

In the experiments of LargeST, we used SD, GBA, GLA and CA datasets with years from 2017 to 2021, followed by LargeST. For example, you can download CA dataset from the provided link and please place the downloaded archive.zip file in the LargeST/data/ca folder and unzip the file.

First of all, you should go through the jupyter notebook process_ca_his.ipynb in the folder LargeST/data/ca to process and generate a cleaned version of the flow data. Then, please go through all the cells in the provided jupyter notebooks generate_SUBDATASET_dataset.ipynb in the folder LargeST/data/SUBDATASET for SUBDATASET=sd, gba, gla, ca. Finally use the commands below to generate the traffic flow data of LargeST for our experiments.

1.2 Generating the TrafficStream Dataset for OOD Setting

We implement extra experiments on TrafficStream. We have prepared the adjacency matrix data for you, you need to unzip the TrafficStream zip data from the provided link and put all the files: 20XX.npz in folder district3F11T17/finaldata into TrafficStream/data.

1.3 Generating the KnowAir Dataset for OOD Setting

You need to download the KnowAir.npy file from the provided link and please place the downloaded Knowair.npy file in the Knowair/data folder to complete the data files.

2. Model Running

To run STOP on LargeST, for example, you may enter the folder LargeST and directly execute the Python file in the terminal:

python experiments/stop/main.py --device cuda:0 --dataset SUBDATASET --checkyears YEARS

for SUBDATASET=sd, gba, gla, ca and YEARS=2017, 2018, 2019, 2020, 2021. To run STOP on Knowair or TrafficStream, you may enter the corresponding folder and directly execute the Python file in the terminal:

python experiments/stop/main.py --device cuda:0

📄 Citation

If you find this project helpful, please cite us:

@inproceedings{ma2025robust,
  title     = {Robust Spatio-Temporal Centralized Interaction for OOD Learning},
  author    = {Jiaming Ma and Bingwu Wang and Pengkun Wang and Zhengyang Zhou and Xu Wang and Yang Wang},
  booktitle = {Proceedings of the Forty-Second International Conference on Machine Learning (ICML)},
  year      = {2025}
}