Sources Localization via Information Back Propagation Mechanism

With the rapid expansion of social networks, therampant spread of rumors and misinformation has increasinglythreatened social stability, making timely and accurate identifi-cation of information sources an urgent need. Traditional sourcelocalization approaches struggle with the inherent randomness ininformation propagation, while existing deep learning methodsoften fail to balance prediction accuracy and model complexityor rely excessively on large-scale observational data. To addressthese challenges, this paper proposes a novel deep learningframework that enables effective Source Localization throughan innovative Information Back Propagation Mechanism (SL-IBPM). Derived from this mechanism, SL-IBPM encapsulatestwo core components: self-information Damping and neighbor-information Aggregation. These components are integrated intothe BackProp module, which can be embedded into variousunidirectional propagation models. This design achieves a fa-vorable balance between prediction accuracy and computationalefficiency and enhances adaptability to different propagationscenarios. Additionally, a weighted loss function is employedto mitigate the class imbalance issue, where non-source nodesfar outnumber source nodes. Extensive experiments on six real-world networks and two real propagation cascade datasets showthat SL-IBPM outperforms state-of-the-art methods in sourcedetection accuracy. It maintains low model complexity andexhibits robust performance even with limited observational data,highlighting its practical superiority.