Joint Multi-modal Graph Structure and Representation Learning for Fake News Detection

With the development of the Internet and multi-media technologies, fake news spread on the Internet has become a problem that cannot be ignored by the public and the government. Multi-modal fake news detection methods have tried to make use of the powerful representation ability of graph convolutional network in the fake news detection task. However, how to learn optimal graph structure for subsequent graph learning and effectively bridge the inter-modality gap is still a challenge. In this paper, we propose a novel approach called Joint Multi-modal Graph Structure and Representation Learning (JMGSRL) for multi-modal fake news detection. It consists of two main modules, i.e., a multi-modal graph structure learning module and a modality discriminator module. The multi-modal graph structure learning module jointly optimizes the graph structures of multiple modalities based on graph convolutional network by adding the potential edges or reweighting the unreasonable edges. The optimized graph structures are further used for subsequent graph learning. To effectively deal with the modality difference issue, the modality discriminator is designed to learn modality-invariant features, and the network training is guided by the adversarial scheme. Experiments on two public real datasets demonstrate that JMGSRL can outperform the state-of-the-art related multi-modal fake news detection methods.