A MTF-GCN: A MULTIVIEW TEMPORAL FUSION GRAPH CONVOLUTIONAL NETWORK FOR RUMOR DETECTION AND FACT-CHECKING ON SOCIAL MEDIA

Abstract

Rumors and low-reliability sources spread very quickly on social media and can have serious impacts on society. Many existing methods rely heavily on textual content or static propagation structures, thus being limited in capturing the dynamic interaction between information and users over time. In this study, we propose MTF-GCN, a multi-perspective time-based consolidation model for the problem of detecting rumors and verifying information at the document level. Each event is represented by two dynamic graphs: (i) propagation graphs constructed from reply and reshare relationships, and (ii) user graphs based on profile and interaction characteristics. The diffusion process is divided into three time phases of early, intermediate and late, which are then processed by graph convolution layers and synthesized through a compact time-consolidation mechanism. The representations obtained from the propagation branch and the user branch are further combined to perform binary layering. Experiments on Weibo, Pheme and ReCOVery datasets showed that MTF-GCN achieved high accuracy (97.75%, 95.22% and 93.10%, respectively), and was particularly strong in recalling untrusted content. These results suggest that multi-perspective spatio-temporal modeling combined with a simple time-based consolidation mechanism is effective and suitable for early warning and misinformation monitoring systems.