Rumor Detection by Jointly Learning Propagation Patterns and Users’ Personas

Information spreads rapidly on social media these days, regardless of its facticity. Consequently, rumor detection has become a critical task. Existing methods have focused on analyzing the content of source claim or the propagation graph formed by responsive posts. However, few studies delve into generating valid representations for users’ personas, which have been shown of great importance according to the journalism and communication research. In this paper, we propose a rumor detection model named $JoPU$ that jointly learns propagation patterns and users’ personas. Generally, our JoPU model is realized by bi-level graph learning. First, users are characterized by their profiles and historical posts, forming a relation graph. Micro-level graph learning is then employed to generate users’ personas representations. Second, on the information propagation graph, we fuse the text features of responsive posts and the learned users’ personas, and then perform macro-level graph learning. Finally, the network is trained in a multi-task way, incorporating the downstream task of rumor detection and an auxiliary task of user similarity, with adaptive tuning of the two tasks weights. Experimental results demonstrate that our model outperform state-of-the-art models.