Official repository for the paper:
"STA-GANN: A Valid and Generalizable Spatio-Temporal Kriging Approach"
Accepted at ACM CIKM 2025.

• Introduction
• Key Contributions
• Repository Structure
• Getting Started
• Datasets
• Baselines
• Citation
• Contact

Spatio-temporal tasks often encounter incomplete data due to missing or inaccessible sensors, making spatio-temporal kriging crucial for inferring missing temporal information.

However, existing models face challenges in:

• Capturing dynamic spatial dependencies and temporal shifts,
• Ensuring validity of spatio-temporal patterns,
• Optimizing generalizability to unknown sensors.

To address these issues, we propose Spatio-Temporal Aware Graph Adversarial Neural Network (STA-GANN), a novel GNN-based kriging framework that enhances both validity and generalization of spatio-temporal pattern inference.

• Decoupled Phase Module (DPM): Detects and adjusts timestamp shifts.
• Dynamic Data-Driven Metadata Graph Modeling (D3MGM): Updates spatial relationships using temporal signals and metadata.
• Adversarial Transfer Learning Strategy: Ensures robust generalization to unseen sensors.

Our STA-GANN and STKriging implementations are developed based on BasicTS (Dec 2023 release).
We plan to continuously follow the updates of BasicTS to ensure that our framework remains fully aligned with the official version.

The current repository structure is as follows:

├── data_preparation/   # Dataset processing methods
├── datasets/           # Raw data
├── examples/           # Parameters CFG for each dataset and each methods
├── ── {Method}/        # Corresponding method folder
├── ── ── {Method}_{Datasets}.py # Each method and dataset has a CFG py file.
├── stkriging/         # Main Folder
├── ── arch/           # Kriging algorithm
├── ── data/           # Dataset processing related
├── ── loss/           # Define the loss function, redirected from metrics
├── ── metrics/        # metrics
├── ── runners/        # pipeline
├── ── utils/          # Kriging processing related
└── README.md          # Project documentation

Please install all dependencies via:

pip install -r requirements.txt

Go to examples/run.py, select the method and dataset you need, then run:

python examples/run.py

For spatio-temporal kriging experiments, we split the data along two dimensions:

1. Node Split

   • 7:1:2 ratio
   • 7 parts for known sensors in training
   • 1 part for unknown sensors in validation
   • 2 parts for unknown sensors in test

2. Series Split

   • 7:3 ratio
   • 7 parts for training
   • 3 parts for validation & test

arXiv link: http://arxiv.org/abs/2508.16161

The citation:

@article{li2025sta,
  title={STA-GANN: A Valid and Generalizable Spatio-Temporal Kriging Approach},
  author={Li, Yujie and Shao, Zezhi and Yu, Chengqing and Qian, Tangwen and Zhang, Zhao and Du, Yifan and He, Shaoming and Wang, Fei and Xu, Yongjun},
  journal={arXiv preprint arXiv:2508.16161},
  year={2025}
}

For any issues, please contact: [email protected]