Temporal Enhanced Multimodal Graph Neural Networks for Fake News Detection

Fake news detection is of crucial importance and has received great attention. However, the existing fake news detection methods rarely consider the news release time, which limits the achievable detection performance, especially for detecting the instant fake news clusters that have sudden and aggregated characteristics. To tackle this issue, a temporal enhanced multimodal graph neural networks (TEMGNNs) method is proposed. The multimodal graph with semantic complementary enhancement is developed by feature aggregation of textual information, image information, and external knowledge. Moreover, the associations among different modalities are obtained by using the graph attention networks and the weights of each modality are adaptively learned. Furthermore, the aggregation of news with adjacent time and the same topic to form a temporal news cluster and learning temporal features for fake new detection by using our proposed graph neural networks. Extensive experiments results obtained on two public datasets demonstrate that our proposed method has the best performance compared with the benchmark methods. It is also shown that the exploitation of the temporal information and multimodal information benefits for fake news detection.