Deep learning of contagion dynamics on complex networks.

Area of Science:

• Computational epidemiology
• Network science
• Machine learning

Background:

• Mechanistic models for contagion dynamics often rely on simplifying assumptions, limiting their predictive accuracy and ability to capture complex behaviors.
• Forecasting disease spread on networks remains a significant challenge in epidemiology.

Purpose of the Study:

• To propose a complementary deep learning approach for modeling contagion dynamics on networks.
• To develop a graph neural network architecture that learns local mechanisms from time series data with minimal assumptions.

Main Methods:

• Utilized a graph neural network architecture to learn contagion dynamics directly from time series data.
• Trained and validated the model on various contagion dynamics of increasing complexity.
• Applied the model to real-world COVID-19 outbreak data from Spain.

Main Results:

• The deep learning approach demonstrated accuracy in forecasting contagion dynamics across different complexities.
• The model successfully simulated dynamics on arbitrary network structures, enabling exploration beyond training data.
• Accurate predictions were achieved for the COVID-19 outbreak in Spain.

Conclusions:

• Deep learning provides a powerful and complementary tool for building effective contagion dynamics models on networks.
• This data-driven approach overcomes limitations of traditional mechanistic models.
• The methodology offers new possibilities for understanding and predicting disease spread.