[NeurIPS 2025] How Different from the Past? Spatio-Temporal Time Series Forecasting with Self-Supervised Deviation Learning

• ST-SSDL Framework:

Preprint Link

• OS

  Linux systems (e.g. Ubuntu and CentOS).

• Python

  The code is built based on Python 3.9. You can install required packages using pip:

  pip install -r requirements.txt
  

• Datasets

  All six datasets are already provided. The PEMS-BAY dataset is provided in zip file due to the limitation of file size. You just need to unzip the file in the PEMSBAY folder.

• Run following commands to prepare data:

  python generate_training_data_his_BAY.py
  python generate_training_data_his_LA.py
  python generate_training_data_his_PEMS.py --dataset PEMS04
  python generate_training_data_his_PEMS.py --dataset PEMS07
  python generate_training_data_his_PEMS.py --dataset PEMS08
  python generate_training_data_his_D7.py --dataset PEMSD7M

• Then train the model with following commands:

  cd model_STSSDL
  python train_STSSDL.py --gpu 0 --dataset METRLA
  python train_STSSDL.py --gpu 0 --dataset PEMSBAY
  python train_STSSDL.py --gpu 0 --dataset PEMSD7M
  python train_STSSDL.py --gpu 0 --dataset PEMS04
  python train_STSSDL.py --gpu 0 --dataset PEMS07
  python train_STSSDL.py --gpu 0 --dataset PEMS08
  

  Performance on Spatiotemporal Forecasting Benchmarks