GNNs beat traditional methods for epidemic source detection

A new study from researchers Martin Sterchi, Nathan Brack, and Lorenz Hilfiker systematically reviews and benchmarks Graph Neural Networks (GNNs) for the source detection problem—identifying the origin of an epidemic spreading over a contact network. The paper, published on arXiv, reproduces and tests four representative GNN architectures against traditional methods (like rumor centrality) and MLP-based baselines under controlled conditions.

Surprisingly, the authors initially set out to challenge the notion that GNNs are superior for this task. However, their experiments across multiple network topologies showed that GNNs substantially outperform all other methods. They also examine how detectability evolves over time, performance scaling with training set size, and sensitivity to observation timing and epidemic parameter uncertainty. The work highlights epidemic source detection as a natural benchmark for evaluating GNN architectures, with all code and data released on GitHub.