Spatio-temporal epidemic forecasting with graph-based transformer.

Area of Science:

• Epidemiology and Public Health
• Artificial Intelligence
• Network Science

Background:

• Accurate epidemic forecasting is crucial for public health, as highlighted by the COVID-19 pandemic.
• Existing spatio-temporal Graph Neural Network (GNN) models struggle with complex temporal dynamics and long-range spatial dependencies in human mobility networks.

Purpose of the Study:

• To develop novel spatio-temporal architectures combining GNNs with Transformer-based temporal modeling.
• To address limitations in capturing non-linear temporal dynamics and long-range spatial dependencies in epidemic forecasting.

Main Methods:

• Introduced two architectures: a local-attention-model and a global-attention-model, integrating GNNs with Transformer temporal modeling.
• Benchmarked against Persistence, Graph Convolutional Recurrent Network (GCRN), and Graph WaveNet using real-world datasets from Spain and Brazil.

Main Results:

• The Linear Temporal Graph Convolutional Network (LinearTGCN) variant achieved superior performance on the Spain dataset (SMAPE 24.74%, MDA 72.52%).
• The local-attention-model performed competitively on the Brazil dataset (RMSE 3873.63, SMAPE 83.47%).
• Simple linear models matched or exceeded Transformer performance on structured data, while Transformers excelled on noisy/unstructured data.

Conclusions:

• Novel spatio-temporal GNNs with Transformer-based temporal modeling offer improved epidemic forecasting.
• Linear models can be highly effective for structured time-series data in this domain.
• Transformer-based models demonstrate strong performance on noisy or unstructured datasets, highlighting their adaptability.