DST-SGNN

September 27, 2025 · View on GitHub

The full paper and appendix are available at Arxiv.

Appendix

The appendix of the paper can be accessed through Appendix. This appendix contains several supplementary materials, such as detailed verification processes, experimental data, supplementary analyses, or other information that supports the main content of the paper but is not included in the main text due to space limitations.

Overview

DST-SGNN: Dynamic Spatio-Temporal Stiefel Graph Neural Network

This project introduces DST-SGNN, designed to efficiently forecast spatio-temporal time series (STTS) data. STTS are widely used in various applications, but their forecasting is challenging due to complex dependencies in both time and space dimensions.

LAD-SGNN: LLM-Augmented Stiefel Graph Neural Networks

LAD-SGNN enhances DST-SGNN by integrating a lightweight large language model with structured prompts and a spatio-temporal alignment mechanism for unified structural and semantic modeling.

Key highlights:

  1. Model design incorporates a thorough reorganization of spatiotemporal forecasting studies, including recent LLM-based approaches.

  2. LLM integration with structured prompts.

  3. Supports zero-shot and few-shot forecasting.

(Note: Files with _L in their names and the scripts RUN_FEW.py and RUN_ZERO.py are related to LAD-SGNN. Specifically, RUN_FEW.py and RUN_ZERO.py are used for few-shot and zero-shot experiments, and their usage is similar to DST-SGNN.)

Prerequisites

  • Python 3.10.14
  • PyTorch 2.1.2
  • torchvision 0.16.2
  • torchaudio 2.1.2
  • PyTorch-CUDA 12.1

Setup Instructions

  1. Install Python 3.10.14 (if you haven't already).

  2. Create and activate a new conda environment (optional but recommended):

    conda create -n your_env_name python=3.10.14
conda activate your_env_name

  3. Install the required libraries: Run the following command to install the necessary dependencies:

    conda install pytorch==2.1.2 torchvision==0.16.2 torchaudio==2.1.2 pytorch-cuda=12.1 -c pytorch -c nvidia

Usage

  1. Place your dataset in the data folder.
  2. Run the RUN.py script: Open the terminal in VSCode and run the following command: 
    python RUN.py

Hyperparameters

The hyperparameters are set via command-line arguments or the hyperparameter.txt file.

Parameters:

  • --data: Dataset (e.g., PEMS03)
  • --feature_size: Feature size (e.g., 358)
  • --batch_size: Batch size (e.g., 2)
  • --train_epochs: Epochs (e.g., 25)
  • --seq_length: Sequence length (e.g., 336)
  • --pre_length: Forecast length (e.g., 96)

Usage:

You can pass parameters directly:

python RUN.py --data PEMS03 --feature_size 358 --batch_size 2 --train_epochs 25 --seq_length 336 --pre_length 96

Or： Modify RUN.py, then run:

python RUN.py