ASTGI: Adaptive Spatio-Temporal Graph Interactions for Irregular Multivariate Time Series Forecasting

🚀 Introduction

Irregular multivariate time series (IMTS) are prevalent in critical domains like healthcare and finance, where accurate forecasting is vital for proactive decision-making.
However, the asynchronous sampling and irregular intervals inherent to IMTS pose two core challenges:

1. Representation: How to accurately represent the raw information of irregular time series without distortion.
2. Dependency Modeling: How to effectively capture complex dynamic dependencies between observation points.

To address these challenges, we propose ASTGI (Adaptive Spatio-Temporal Graph Interaction) — a simple yet powerful framework:

• 🧩 Point-wise Representation: Encode each observation as a spatio-temporal point in a learnable embedding space.
• 🔗 Neighborhood-Adaptive Graph Construction: Dynamically build causal graphs via nearest neighbor search.
• 🔄 Dynamic Propagation: Iteratively update node features through spatio-temporal message passing.
• 🎯 Query-based Prediction: Aggregate neighbors to forecast the value at a new query point.

📦 Quickstart

Recommended Environment

• Python 3.11
• CUDA 12.0

1️⃣ Installation

Create a clean environment and install dependencies:

conda create -n astgi python=3.11 -y
conda activate astgi
pip install -r requirements.txt

2️⃣ Data Preparation

The PhysioNet, USHCN, and Human Activity datasets are downloaded and processed automatically when you run the code for the first time.

For the MIMIC dataset, a manual preprocessing step is required:

1. Follow the preprocessing scripts in gru_ode_bayes to generate the complete_tensor.csv file.
2. Place the generated file under the following path (you may need to create the directories):

   ~/.tsdm/rawdata/MIMIC_III_DeBrouwer2019/complete_tensor.csv

3️⃣ Train

Run the benchmark experiments with provided scripts:

bash ./scripts/ASTGI/USHCN.sh

📊 Benchmark Results

📈 ASTGI consistently outperforms SOTA baselines across multiple datasets.

🤝 Contributing

Contributions are welcome! Feel free to open issues or submit pull requests if you find bugs or want to add new features.