Leveraging Heterophily in Spatial-Temporal Graphs for Multivariate Time-Series Forecasting

Structure:

• PEMS03, PEMS04, PEMS07 and PEMS08 datasets are released by and available at STSGCN. \ KnowAir dataset is released by and available at PM2.5-GNN.

• lib: contains self-defined modules for our work, such as data loading, data pre-process, normalization, and evaluate metrics.

• model: implementation of our DPSTGNN model

Requirements

Python 3.8.11, Pytorch 1.9.0, Numpy 1.12.2, tslearn, argparse and configparser

Here, we present the results of different sliding window size parameters on the PEMS04 and PEMS08 datasets.

Windows Size	1			3			5
	MAE	MAPE	RMSE	MAE	MAPE	RMSE	MAE	MAPE	RMSE
PEMS04	19.11	12.68%	30.99	18.71	12.23%	30.79	19.18	12.70%	18.83
PEMS08	14.83	9.98%	24.34	14.42	9.31%	23.87	14.71	9.88%	24.13