Early Rumour Detection with Temporal Bidirectional Graph Convolutional Networks

Automatic rumour detection has drawn significant research attention and deep learning models are proposed. It is shown that misinformation propagates further and wider on social networks. Existing research has focused on using the information propagation pattern for rumour detection. But the temporal propagation pattern for rumours has been largely ignored. This paper addresses this gap. We propose a temporal Bi-directional Graph Convolutional Network (tBi-GCN) model to learn representations for rumour propagation and rumour dispersion by encoding the temporal information for local graph structures and nodes. Specifically, we constructed a time-weighted adjacency matrix to represent the effect of time delay between nodes on information dissemination. Experimental results across five events of the PHEME dataset show that tBi-GCN can achieve a comparable performance in comparison with several state-of-the-art models for early rumour detection.