Graph Neural Networks for Source Detection: A Review and Benchmark Study

Graph neural networks substantially outperform traditional methods at identifying the origin of an epidemic spreading across a contact network, according to a systematic benchmark study that set out to challenge the technology’s effectiveness [1][2]. The source detection problem, formally defined in 2010 by Shah and Zaman, asks analysts to trace an unfolding epidemic back to its point of origin using only the pattern of observed infections [1][2]. The task has long been approached with classical algorithms, but the rise of graph neural networks — specialized architectures that pass messages between nodes to learn from graph-structured data — prompted multiple research groups to propose GNN-based solutions [2][3]. Until now, inconsistent experimental setups made it difficult to determine whether the newer methods genuinely improved on older ones [2]. Martin Sterchi and collaborators reproduced four representative GNN architectures and tested them against traditional baselines and multi-layer perceptron models under controlled, comparable conditions [1][2]. “Our experiments show that GNNs substantially outperform all other methods we test across a variety of network topologies,” the authors write, noting that they initially expected to challenge the notion of GNNs as a solution [2]. The study also examined how detectability changes over time, how performance scales with training-set size, and how sensitive the models are to uncertainty in observation timing and epidemic parameters [2]. GNNs operate by iteratively updating node representations through pairwise message passing, a design that has proven effective in domains ranging from molecular drug design to natural language processing [3]. The source-detection task shares conceptual ground with anomaly detection, where the goal is to identify rare events that deviate from normal behavior — a problem class that appears in cybersecurity, fraud analysis, and neuroscience [4]. The benchmark results position epidemic source detection as a natural testbed for evaluating GNN architectures, and the research team has released all code and data on GitHub to support reproducibility [2]. The paper was submitted to arXiv in December 2025 and revised in May 2026 [1].