michaelajao/MSAGAT-Net: MSAGAT-Net

MSAGAT-Net MSAGAT-Net is a deep learning framework designed to address three fundamental challenges in epidemic forecasting: computational scalability, multi-scale temporal dynamics, 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