STA-GANN: A Valid and Generalizable Spatio-Temporal Kriging Approach

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

📖 Table of Contents

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

---

📖 Introduction

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.

---

🔑 Key Contributions

• 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.

---

📂 Repository Structure

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

🚀 Getting Started

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

---

📊 Datasets

Dataset	Domain	Sensors	Duration	Time Steps	Frequency	Extra Info	Data Link
METR-LA	Traffic	207	Mar 2012 – Jun 2012	34,272	5 min	Includes latitude/longitude; sensor graph; adjacency via Gaussian kernel	LINK
PEMS-BAY	Traffic	325	Jan 2017 – Jun 2017	52,116	5 min	Includes latitude/longitude; sensor graph; adjacency via Gaussian kernel	LINK
PEMS03	Traffic	358	Sep 2018 – Nov 2018	26,208	5 min	Sensor graph only; no latitude/longitude	LINK
PEMS04	Traffic	307	Jan 2018 – Feb 2018	16,992	5 min	Sensor graph only; no latitude/longitude	LINK
PEMS07	Traffic	883	May 2017 – Aug 2017	28,224	5 min	Sensor graph only; no latitude/longitude	LINK
PEMS08	Traffic	170	Jul 2016 – Aug 2016	17,856	5 min	Sensor graph only; no latitude/longitude	LINK
NREL	Energy	137	2006	105,120	10 min*	Solar power plants in Alabama; includes latitude/longitude	LINK
USHCN	Climate	1,218	1899 – 2019	1,440	Monthly	Precipitation; includes latitude/longitude	LINK
AQI	Environment	437	43 cities (China)	59,710	Hourly*	Air Quality Index (PM2.5); includes latitude/longitude	LINK

📌 Data Split

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

---

📚 Baselines

Name	Paper Title	Venue	Year	Link	Type
GCN	Semi-Supervised Classification with Graph Convolutional Networks	ICLR	2017	LINK	Backbone
GIN	How Powerful are Graph Neural Networks?	ICLR	2019	LINK	Backbone
IGNNK	Inductive Graph Neural Networks for Spatiotemporal Kriging	AAAI	2021	LINK	Spatio-temporal Kriging
GRIN	Filling the Gaps: Multivariate Time Series Imputation by Graph Neural Networks	ICLR	2022	LINK	Adapted Spatio-temporal Imputation
SATCN	Spatial Aggregation and Temporal Convolution Networks for Real-time Kriging	None	2021	LINK	Spatio-temporal Kriging
INCREASE	INCREASE: Inductive Graph Representation Learning for Spatio-Temporal Kriging	WWW	2023	LINK	Spatio-temporal Kriging
DualSTN	Decoupling Long- and Short-Term Patterns in Spatiotemporal Inference	TNNLS	2023	LINK	Spatio-temporal Kriging
IAGCN	Inductive and Adaptive Graph Convolution Networks Equipped with Constraint Task for Spatial–Temporal Traffic Data Kriging	KBS	2024	LINK	Spatio-temporal Kriging
OKriging	—	—	—	—	Traditional Kriging

---

📜 Citation

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

The citation:

@inproceedings{li2025sta,
  title={STA-GANN: A Valid and Generalizable Spatio-Temporal Kriging Approach},
  author={Li, Yujie and Zezhi, Shao and Yu, Chengqing and Qian, Tangwen and Zhang, Zhao and Du, Yifan and He, Shaoming and Wang, Fei and Xu, Yongjun},
  booktitle={Proceedings of the 34th ACM International Conference on Information and Knowledge Management},
  pages={1726--1736},
  year={2025}
}

---

🤝 Contact

For any issues, please contact: liyujie23s@ict.ac.cn