Observer-free source detection in temporal networks with graph neural networks

Effective source detection in temporal networks is critical for managing epidemics, yet is hampered by prevailing methods that assume static network structures, uniform outbreak times, or rely on observer nodes. To overcome these limitations, we propose a novel graph neural network method, Node Heterogeneity and Temporal Dynamics (NHTD), designed for robust, observer-free source tracing under realistic conditions. NHTD synergistically integrates three feature classes capturing node local heterogeneity, multi-timescale influence, and time-information flow potential, unified by an attention mechanism to prioritize salient information. Operating with only the final infection states and network dynamics, our model adeptly handles varying and unknown outbreak times. Extensive experiments on diverse real-world temporal networks demonstrate that NHTD significantly outperforms state-of-the-art baselines in both accuracy and robustness, particularly in complex and noisy environments. Our work provides a scalable and practical solution for source detection, offering a significant advancement for real-time epidemiological surveillance.