UVA's 'boundary degree' feature boosts epidemic simulation accuracy by 19%

A team from the University of Virginia's Biocomplexity Institute has introduced a novel node-level feature called 'boundary degree' to improve how AI models identify epidemic scenarios from simulated disease cascades. The feature is defined as the number of an infected node's contacts in the underlying social network that remain uninfected. Through systematic ablation on realistic contact networks from Tennessee and Virginia, the researchers showed that adding boundary degree alone boosts scenario identification accuracy by 19%.

Edge features, which were previously shown to be important empirically, also consistently improved accuracy across all settings—and this work provides the first theoretical grounding for that observation. Crucially, the effects of boundary degree and edge features are complementary. The authors prove that without either boundary or edge information, certain epidemic scenarios become mathematically indistinguishable, making these features essential for accurate classification.

Prior approaches had included aggregate boundary statistics, but these were never among the top-ranked feature groups. The per-node representation used here reveals their importance clearly. The findings carry a practical implication: contact tracing applications should consider tracking contacts with non-infected individuals, not only confirmed transmission chains. The paper is published as arXiv:2606.29596 and includes 28 pages with 10 figures.