Multi-source localization in social networks via hybrid sampling and Bayesian-inspired graph optimization

Information diffusion source localization is critical in complex social networks. However, existing source localization methods fail to sufficiently address the ill-posed inverse mapping between information spreaders and observed network snapshots, especially in large-scale networks. To tackle these challenges, we introduce a multi-source localization approach based on Hybrid Sampling and Bayesian-inspired Graph Optimization(HSBGO). Specifically, inspired by the Bayesian optimization, we develop a graph optimization module to capture the stochastic ambiguity in the source–snapshot mapping by employing a GCN-based surrogate model. Furthermore, we introduce a hybrid sampling strategy to alleviate the search space inflation in multi-source localization by integrating the coarse sampling and the feature aggregated refinement. It is worth mentioning that HSBGO can handle various diffusion models and does not require prior information about the number of sources. Extensive experiments on multiple real-world and synthetic networks demonstrate that our approach consistently outperforms state-of-the-art methods.