4HAN: hypergraph-based hierarchical attention network for fake news prediction

Abstract

Fake News presents significant threats to both society and individuals, highlighting the urgent need for improved news authenticity verification. To deal with this challenge, we provide a novel strategy called the 4-level hierarchical attention network (4HAN), designed to enhance fake news detection through an advanced integration of hypergraph convolution and attention neural network mechanisms. The 4HAN model operates across four hierarchical levels: paragraphs, sentences, words, and contextual information (metadata). At the highest level, the model employs hypergraph-based attention and convolution neural networks to create a contextual information vector, utilizing a SoftMax activation function. This vector is then combined with a news content vector generated through word and sentence-level attention mechanisms. This architecture enables the 4HAN model to effectively prioritize the relevance of specific words and contextual information, thereby improving the overall representation and accuracy of news content. We evaluate the 4HAN model using the LIAR dataset to demonstrate its efficacy in enhancing Fake News prediction accuracy. Comparative analysis shows that the 4HAN model outperforms several of cutting-edge techniques, like recurrent neural networks (RNN), ensemble techniques, and attention mechanisms techniques. Our results indicate 4HAN model accomplishes a notable accuracy of 96%, showcasing its potential for significantly advancing fake news prediction.