Multiple rumor source localization in signed social networks driven by data intelligence

With the help of deepfake algorithms and social bots, the problems of consciousness penetration and cognitive manipulation caused by the outbreak of internet rumors have become prominent. Accurately locating the rumor sources and quickly cutting off the critical paths of rumor propagation will become an effective ways to curb the explosive spread of rumors and the sudden accumulation of negative emotions. In this paper, we combine the complementary advantages of signed graph convolutional networks in spatial and spectral domains, propose an improved multirumor source localization framework for signed social networks, and extend the rumor centrality principle and source prominence theory to the scope of signed networks. First, structural balance theory is used to accurately model positive and negative social relations. Second, a signed graph convolutional network based on signed attention mechanism is proposed to extract the rumor centrality feature from the infection subgraph. Then, a two-stream graph convolutional network based on a label propagation mechanism is proposed to extract the source prominence feature from the subgraph containing only positive edges and the subgraph containing only negative edges, respectively. Finally, the center feature of the infected structure and the position distribution feature of infected and uninfected nodes are integrated in a unified framework for multirumor source localization. Extensive experimental results on four real-world social network datasets show that compared with state-of-the-art algorithms, our proposed algorithm further improves the accuracy and robustness in the task of multiple rumor source localization.