MSAGAT-Net

MSAGAT-Net

MSAGAT-Net is a deep learning framework designed to address three fundamental challenges in epidemic forecasting: computational scalability, multi-scale spatial dependencies, and multi-horizon forecast stability. The framework achieves O(N) linear complexity through linearised attention with low-rank projections, enabling real-time surveillance across large-scale regional networks.

Architecture Components

• Efficient Feature Extraction: Depthwise separable convolutions with low-rank bottleneck projections for parameter-efficient temporal encoding
• Adaptive Graph Attention Module (AGAM): Linear O(N) complexity attention mechanism with learnable graph structure bias and multi-head attention
• Multi-Scale Spatial Feature Module (MSSFM): Parallel dilated convolutions at multiple scales (dilation rates: 1, 2, 4, 8) with adaptive fusion
• Progressive Prediction Refinement Module (PPRM): Stabilises multi-horizon forecasts through adaptive blending of model predictions with trend extrapolation

Datasets

This release includes seven epidemic forecasting benchmarks:

• Japan Prefecture-level Influenza
• US Regional Data
• US State-level Data
• Australia COVID-19
• Spain COVID-19
• UK LTLA COVID-19
• NHS England ICU Bed Occupancy