Tracing Epidemic Source of Dynamic Network Based on Graph Neural Network

Identifying the source of an epidemic involves inferring its origin based on observations of the infected population, where timely isolation of the source is critical to mitigating outbreak spread. Existing approaches predominantly assume static network topologies, yet real-world networks exhibit dynamic changes in nodes and edges over time, significantly challenging conventional source detection methods designed for static networks. To address this limitation, we propose a dynamic network source detection framework based on a <underline xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">M</u>ulti-observation <underline xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">S</u>ource <underline xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">F</u>usion <underline xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">S</u>patio-<underline xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">T</u>emporal <underline xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">G</u>raph <underline xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">C</u>onvolutional <underline xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">N</u>etwork (MSF-STGCN). Our method first partitions epidemic data into multi-frame static graphs, distinguishing between observed and unobserved node states relative to the target detection time. Next, we leverage graph neural networks to extract spatio-temporal features from multi-snapshot sequence data. Finally, we introduce a dynamic message-passing algorithm to process unobserved network components and integrate the results with multi-snapshot spatio-temporal features for accurate source identification. Extensive experiments on synthetic networks and real-world datasets demonstrate the superior performance of our approach compared to state-of-the-art baselines.