文献类型：期刊文献

英文题名：Source identification in social networks based on graph transformer

作者：Lu, Yisong[1];Wei, Pengfei[1];Li, Jinsheng[1];Li, Qiong[2];Zhang, Xueqin[1]

机构：[1]East China Univ Sci & Technol, Sch Informat Sci & Engn, 130 Meilong Rd, Shanghai 200237, Peoples R China;[2]East China Univ Sci & Technol, Sch Social & Publ Adm, 130 Meilong Rd, Shanghai 200237, Peoples R China

年份：2025

卷号：200

外文期刊名：CHAOS SOLITONS & FRACTALS

收录：;EI(收录号：20253218934841);WOS:【SCI-EXPANDED(收录号:WOS:001546950100002)】；

基金：This work was supported by the Major Program of National Fund of Philosophy and Social Science of China (grant number: 23&ZD142) .

语种：英文

外文关键词：Social networks; Source identification; Graph transformer; Label propagation; Imbalanced classification

摘要：With the rapid expansion of social network scale and the exponential growth of information diffusion speed, tracing the source of misinformation has become a critical task for ensuring the security of the online ecosystem. In recent years, Graph Neural Networks have demonstrated commendable performance in source node detection. However, their neighborhood aggregation mechanisms suffer from over-smoothing and are restricted to local neighborhood modeling, rendering them incapable of capturing long-range dependencies, and fully exploiting global traceability information. To address these challenges, we propose an end-to-end source identification framework, named GTSI (Graph Transformer Source Identification). This method is based on the graph attention mechanism, aiming to overcome the dependency on specific propagation models and local perception limitations of existing methods. GTSI introduces a graph attention bias module based on diffusion labels to adaptively model implicit propagation associations between nodes. It integrates static topological features with dynamic infection states to construct a heterogeneous node representation space. Additionally, a composite loss function is designed to mitigate prediction biases arising from class imbalance. Experimental results demonstrate that, compared to the baseline methods, GTSI provides a novel technical pathway for information traceability tasks, exhibiting superior identification accuracy and cross-scenario transferability on real-world social networks and diverse propagation models.