STSGT

STSGT forecasts COVID-19 trends by capturing complex spatial and temporal dependencies in pandemic time series data.

Key Features:

• Integration of Graph Convolution Network (GCN) and Transformers: Combines GCNs for spatial analysis with transformer self-attention for temporal dynamics on synchronous spatial-temporal graphs.
• Synchronous Spatial-Temporal Graph: Constructs a graph that simultaneously represents spatial relationships (e.g., between geographic regions or states) and temporal dependencies (e.g., daily case sequences).
• Enhanced Forecasting Accuracy: Demonstrates reduced Mean Absolute Error (MAE) for predicting daily infected and death cases by capturing both spatial and temporal patterns.

Scientific Applications:

• Epidemiological forecasting: Forecasts COVID-19 trends across geographic regions, including state-wide and county-specific analyses within the United States.
• Case and mortality prediction: Predicts daily infected and death counts using spatiotemporal dependencies in time series data.
• Model validation on public datasets: Validated on two publicly available COVID-19 time series datasets to assess forecasting performance.

Methodology:

Constructs a synchronous spatial-temporal graph representing spatial connections (e.g., between states) and temporal sequences (daily case numbers), processes spatial information with GCN layers and temporal dynamics with transformer layers to adapt to changing patterns in the data.