RESEAT: Recurrent Self-Attention Network for Multi-Regional Influenza Forecasting.

Area of Science:

• Epidemiology
• Computer Science
• Data Science

Background:

• Early forecasting of influenza is vital for public health interventions and minimizing societal losses.
• Existing deep learning models for multi-regional forecasting often struggle to jointly capture complex temporal and regional patterns.
• Current attention-based models have limitations in dynamically modeling evolving regional interrelationships.

Purpose of the Study:

• To propose a novel deep learning model, the recurrent self-attention network (RESEAT), for enhanced multi-regional forecasting.
• To address the limitations of existing models in capturing dynamic regional interdependencies over time.
• To improve the accuracy of forecasting tasks such as influenza and electrical load prediction.

Main Methods:

• Development of the recurrent self-attention network (RESEAT) architecture.
• Utilizing self-attention to learn regional interrelationships across the entire input data period.
• Employing message passing to recurrently connect attention weights, enabling dynamic modeling of interrelationships.

Main Results:

• The proposed RESEAT model demonstrates superior forecasting accuracy compared to state-of-the-art models.
• Experimental validation shows significant improvements in predicting influenza and COVID-19 outbreaks.
• The study includes methods for visualizing regional interrelationships and analyzing hyperparameter sensitivity.

Conclusions:

• RESEAT offers a significant advancement in multi-regional forecasting by effectively modeling dynamic regional patterns.
• The model's ability to capture evolving interrelationships leads to improved accuracy in public health and other forecasting applications.
• The findings provide a foundation for more robust and accurate epidemiological and load forecasting systems.