Long-term Spatio-Temporal Forecasting via Dynamic Multiple-Graph Attention

Introduction

We propose a new dynamic multi-graph fusion module to characterize the correlations of nodes within a graph and the nodes across graphs via the spatial attention and graph attention mechanisms. Furthermore, we introduce a trainable weight tensor to indicate the importance of each node in different graphs. We can increase the performance of plenty of STGCNN models including STGCN, ASTGCN, MSTGCN, ST-MGCN, and Graph WaveNet.

Framework

Datasets

Histocical Records

• Air Quality The Ministry of Ecology and Environment of China (MEE) published a large-scale air quality dataset, comprising 92 air quality monitoring stations, to assess the hourly PM2.5 concentration in Jiangsu province in 2020.

Weight Matrix

There are five weight matrices:

• Distance Graph:
• Neighbor Graph:
• Functionality graph:
• Heuristic Graph:
• Temporal Pattern Similarity Graph:

Multi-graph Spatial Embedding

The multi-graph spatial embedding is generated using node2vec.

The historical records are stored in 'data\temporal_data', the weight matrix data is stores in 'data\graph', and the multi-graph spatial embedding data is stores in 'data\SE'.

Requirements

Our code is based on Python3 (>= 3.6). The major libraries are listed as follows:

• torch >= 1.8.1
• numpy >= 1.15
• scipy >= 1.1.0
• torch-cluster >= 1.5.9
• torch-geometric >= 1.7.2
• torch-scatter >= 2.0.6
• torch-sparse >= 0.6.9
• torch-spline-conv >= 1.2.1
• pytorch-lightning >= 1.2.8
• wandb >= 0.11.1

The following command can help install the above libraries:

pip install -r requirements.txt

Usage

1. Generate training data The time steps of historical observations and prediction horizons are both set to 24. The train, validation, and test part are divided into 7:1:2. Run the following command to generate training data:

python generate_training_data.py

2. Parameters

• graphs: distance graph, neighbor graph, functionality graph, heuristic graph, temporal pattern similarity graph
• model: ASTGCN
• length of histotical observations: 24
• length of prediction horizons: 24
• epoches: 40
• batch size: 32
• learning rate: 1e-4
• weight decay: 1e-4
• number of attention heads: 8
• dimension of each attention outputs: 8

3. Train the model Run the following command to train model:

python train.py

Citation

Please refer to our paper. Wei Shao*, Zhiling Jin*, Shuo Wang, et al, Flora Salim, Long-term Spatio-Temporal Forecasting via Dynamic Multiple-Graph Attention. Long-term Spatio-Temporal Forecasting via Dynamic Multiple-Graph Attention. In Proceedings of the 31st International Joint Conference on Artificial Intelligence and the 25th European Conference on Artificial Intelligence (IJCAI-ECAI 2022), 2022.

@inproceedings{shao2022long,
  title={Long-term Spatio-Temporal Forecasting via Dynamic Multiple-Graph Attention},
  author={Shao, Wei and Jin, Zhiling and Wang, Shuo and Kang, Yufan and Xiao, Xiao and Menouar, Hamid and Zhang, Zhaofeng and Zhang, Junshan and Salim, Flora},
  booktitle={31st International Joint Conference on Artificial Intelligence, IJCAI 2022},
  pages={2225--2232},
  year={2022},
  organization={International Joint Conferences on Artificial Intelligence}
}