Heterogeneous Propagation Graph Representation Learning for Fake News Detection

Recently, social network platforms have led to a deluge of fake news, which tends to mislead readers and may arouse public spark panic. Thus, detecting fake news on social networks re-mains an important role. Previous researches mainly focused on discovering features from diffusion paths and news content. However, these researches insufficiently exploit dynamic interactions on social networks to refine representations of social roles. Therefore, we proposed hetero-geneous propagation graph representation learning (HPGR), which exploits heterogeneous interaction behavior and non-Euclidean propagation preferences of news to improve the detection performance of fake news. Different from previous works, the HPGR transfers spatial propagation activities into a heterogeneous propagation graph. Centered on the news, the graph iteratively traces its publisher representation and user representation to fully learn the static and dynamic representations of nodes. Specifically, we develop a publisher-based attention module to learn publisher representation in the news propagation network, which can capture the shared semantics of publisher's news more accurately and facilitate the fine-grained representation. Then, the fine-grained user representations are constructed by user similarity and pref-erences in the news propagation. Further, the representations of news contents are reconstructed through propagation and aggregation of propagation graph neural network, which inter-pretively describes the feedbacks of news propagation process. Finally, the reconstructed representations of news contents are applied for fake news detection task. The experimental results show that the HPGR can improve the accuracy of fake news detection and provide a novel paradigm for fine-grained heterogeneous graph representation learning.