Temporal-spatial hierarchical contrastive learning for misinformation detection: A public-behavior perspective

Abstract

The widespread dissemination of misinformation on social media platforms significantly affects public security. Current methods for detecting misinformation predominantly rely on semantic information and social context features. However, they often neglect the intricate noise issues and unreliable information interactions resulting from diverse public behaviors, such as cognitive biases, user prejudices, and bot activity. To tackle these challenges, we propose an approach named TSHCL (temporal-spatial hierarchical contrastive learning) for automatic misinformation detection from the public-behavior perspective. First, the integration of a graph convolutional network (GCN)-based autoencoder architecture with a hybrid augmentation method is designed to model typical public behaviors. Next, node-level contrastive learning is designed to maintain the heterogeneity of comments in the spatial view under the influence of complex public behaviors. Finally, cross-view graph-level contrastive learning is designed to promote collaborative learning between the temporal sequence view of events and the spatial propagation structure view. By conducting temporal-spatial hierarchical contrastive learning, the model effectively retains crucial node information and facilitates the interaction of temporal-spatial information. Extensive experiments conducted on real datasets from MCFEND and Weibo demonstrate that our model surpasses the state-of-the-art models. Our proposed model can effectively alleviate the noise and unreliable information interaction caused by public behavior, and enrich the research perspective of misinformation detection.