Joint learning of structural and textual information on propagation network by graph attention networks for rumor detection

Due to the advantages in information dissemination, social media is growing rapidly among the public but has also become a medium for the spread of rumors. Given the serious damage rumors bring to society, detecting rumors from the mass information on social media is becoming an arduous challenge. Therefore, some deep learning techniques, such as Recurrent Neural Networks and Graph Neural Networks, have been applied to detect rumors with propagation paths and content. Nonetheless, these deep learning-based methods ignore the propagation structure of rumors or cannot effectively extract the propagation structure information of each node. Inspired by this, we propose a novel rumor detection framework named \(\varvec{Rumor2Topic}\) $\mathit{R}\mathit{u}\mathit{m}\mathit{o}\mathit{r}\mathbf{2}\mathit{T}\mathit{o}\mathit{p}\mathit{i}\mathit{c}$ . We first adopt \(\varvec{GraphAnchorLDA}\) $\mathit{GraphAnchorLDA}$ (Graph Anchor Latent Dirichlet Allocation) to extract the structural topic representation of rumor events, and then utilize Graph Attention Networks (GATs) to further extract the complex semantic and propagation structure information of rumor events. Meanwhile, the encouraging empirical results of \(\varvec{Rumor2Topic}\) $\mathit{R}\mathit{u}\mathit{m}\mathit{o}\mathit{r}\mathbf{2}\mathit{T}\mathit{o}\mathit{p}\mathit{i}\mathit{c}$ on several benchmarks confirm that the proposed method is superior to the most advanced methods. Our proposed Rumor2Topic achieved 2.62%, 2.07%, and 2.57% improvements compared to the optimal method on three publicly available benchmarks, Twitter15, Twitter16, and Weibo, respectively.