Knowledge Graph-Augmented Heterogeneous Graph Neural Networks for Fake News Detection

With the development of telecommunication and social networks, millions of information and messages now exist on websites. However, these vast tracts of information vary in quality given confirmation bias, the influence of algorithms, limited expertise, malicious intent, and that there are no regulations over how to process data. Additionally, different people publish messages on social media for a range of reasons – some good-intentioned and some malevolent. Malevolent users can create and disseminate low-quality information for illegal purposes, as a wilful attempt to mislead or deceive, or simply to ‘troll’ people. This inferior information can have a serious impact on people as browsing social networks is one of the main ways many people get their news these days. As such, the ability to detect fake news has become increasingly vital. Knowledge graphs (KGs) contain numerous valuable and ground-truth information about real-world entities that appear in published news or messages, which can significantly contribute to detecting fake news. However, most existing research only focuses on extracting linguistic or semantic information to identify whether news is either fake or real; it fails to fully utilise the ground knowledge in KGs. In this thesis, we aim to develop effective fake news detection methods based on KGs and graph neural networks (GNNs) to fill these gaps. To explore the impact of ground-truth knowledge in KGs on identifying fake news, we first propose a novel KG-enhanced heterogeneous GNN for fake news detection. Then, we propose a heterogeneous GNN via knowledge relations for fake news detection by considering the complex relations between knowledge in a KG and news attributes. In addition, complementary knowledge between different KGs can provide a more comprehensive factual basis for news with large amounts of data. Therefore, we propose a contrastive multi-KG learning for fake news detection. Furthermore, large language models (LLM) have achieved promising performance for text embeddings by incorporating huge amounts of texts in a pre-training regimen. Accordingly, we propose a multi-KG and LLM-inspired heterogeneous GNN for fake news detection. Through extensive experiments on four fake news datasets, we demonstrate the superiority of our devised models over different baseline methods in terms of the most commonly used evaluation metrics, including accuracy, precision, recall, and F1-score.<p></p>