融合时间流特征和传播结构特征的谣言检测

Rumor Detection Integrating Temporal Flow and Propagation Structure

现存关于谣言检测的研究方法要么只关注谣言在社交媒体上传播的时间流特征,要么仅关注传播结构特征,并且使用了大量的辅助信息。实际上,谣言传播的时间流和传播结构特征均有助于提升谣言检测模型的性能,并且能够形成互补作用。与此同时,源用户的自我描述相比于其他辅助信息更为重要,并且源推文的语义信息在整个会话线程中起到了关键作用。为解决上述问题,该文提出了一个新颖的谣言检测模型TPSS。该模型融合了时间流和传播结构特征。同时,仅采用源用户的自我描述作为辅助信息,并且提出了一种协同注意力机制来增强源推文的作用。该机制基于源推文特征来增强时间流特征和传播结构特征。在Twitter15、Twitter16和PHEME数据集上的实验结果表明TPSS优于基准系统。

Existing methods for the rumor detection(RD)considered either only temporal flow or propagation structure in social media,and used numerous auxiliary information.Actually,both the temporal flow and propagation structure are helpful to RD and can complement each other.Moreover,the self-description of source user is more important than other auxiliary information and semantic information in source post.To address the above issues,this paper proposes a novel RD model named TPSS,which integrates the temporal flow based method and the propagation structure based method.Moreover,it only employs the self-description of source user as auxiliary feature and proposes a co-attention mechanism to enhance temporal flow and propagation structure based source post.The experimental results on Twitter15,Twitter16 and PHEME show that TPSS outperforms several state-of-the-art baselines.