Rumor Detection with Field of Linear and Non-Linear Propagation

The propagation of rumors is a complex and varied phenomenon. In the process of rumor dissemination, in addition to rumor claims, there will be abundant social context information surrounding the rumor. Therefore, it is vital to learn the characteristics of rumors in terms of both the linear temporal sequence and the non-linear diffusion structure simultaneously. However, in some existing research, time-dependent and diffusion-related information has not been fully utilized. Accordingly, in this paper, we propose a novel model Rumor Detection with Field of Linear and Non-Linear Propagation (RDLNP) to automatically detect rumors from the above two fields by taking advantage of claim content, social context and temporal information. First, the Rumor Hybrid Feature Learning (RHFL) we designed can extract the correlations between the claims and temporal information, differentiate the hybrid features of specific posts, and generate unified representations for rumors. Second, we proposed Non-Linear Structure Learning (NLSL) and Linear Sequence Learning (LSL) to integrate contextual features along the path of the diffusion structure and temporal engagement variation of responses respectively. Finally, Shared Feature Learning (SFL) models the representation reinforcement and learns the mutual influence between NLSL and LSL, and then highlights their valuable features. Experiments conduct on two public and widely used datasets, i.e. PHEME and RumorEval, demonstrate both the effectiveness and the outstanding performance of the proposed approach.