ForecastGrapher

Paper Link: ForecastGrapher: Redefining Multivariate Time Series Forecasting with Graph Neural Networks

Training Scripts

• Place the datasets in the ./dataset/ folder (for example，./dataset/ETTm1.csv, ./dataset/PEMS03/PEMS03.npz).

• Run the sh file in the ./scripts folder , using the following command: sh ./scripts/ForecastGrapher_***.sh.

Model

Overall Structure

1. Consider each time series as a node and generate a corresponding node embedding.
2. Employ learnable scalers to partition the node embedding into multiple groups.
3. Several layers of GFC-GNN are stacked.
4. Utilize node projection for forecasting.

Learnable Scaler and Group Feature Convolution

An automatic adjustment of feature distributions can be achieved within CNNs. The GFC mechanism enhances the diversity of node embedding distributions: Convoluting the node feature with two distinct kernel lengths results in two distinct distributions.

Multivariate Time Series Forecasting Results

• Comparison with Benchmarks (Avg)

  

• Comparison with advanced GNNs and Naive Method (Avg)

  

Acknowledgement

We appreciate the valuable contributions of the following GitHub.

• iTransformer (https://github.com/thuml/iTransformer)
• LTSF-Linear (https://github.com/cure-lab/LTSF-Linear)
• Time-Series-Library (https://github.com/thuml/Time-Series-Library)
• FourierGNN (https://github.com/aikunyi/FourierGNN)
• StemGNN (https://github.com/microsoft/StemGNN)