Deep learning for fake news detection on social networks

<p dir="ltr">The rise of online social networks has dramatically reshaped how information is shared and exchanged worldwide. Social network platforms such as Twitter, Facebook, Instagram, and TikTok provide unprecedented opportunities for individuals to share, discuss, and exchange information at any time from any location. Yet, the expansive reach and influence of these social networks also introduce substantial challenges, notably the pervasive issue of fake news. The same attributes that make social media a powerful medium for information dissemination also make it prone to the spread of fake news. In response to the critical threats posed by fake news on social networks, this thesis aims to design novel deep learning models for detecting fake news on social networks, tackling the intricate and continuously changing problem of fake news. Initially, we propose a novel framework that utilizes a topic knowledge graph for analyzing data on social networks. This approach allows for a detailed examination of information shared on social networks, both genuine and false, revealing their components, fine-grained characteristics, and semantic relationships. Such insights are invaluable for guiding the design of fake news detection models. Subsequently, we address the emotional biases and knowledge inaccuracies that are often presented in fake news. To this end, we propose the EmoKnow, an emotion- and knowledge-oriented model for identifying fake news. Given the limitations imposed by the unavailability or computational expense of auxiliary information, we then focus on achieving robust fake news detection using solely textual inputs. We propose the heterogeneous subgraph transformer (HeteroSGT), a novel method that investigates semantic patterns at both word and sentence levels, as well as the structural relationships among news content, entities, and topics. Further enhancing our methodology, we propose LESS4FD, an LLM-enhanced semantic mining model designed to harness the capabilities of Large Language Models (LLMs). LESS4FD leverages LLMs as feature enhancers, capitalizing on both local and global semantics and utilizing training signals from unlabeled data to boost detection effectiveness. Through extensive experiments conducted on widely used public fake news datasets, our results validate the superiority of the proposed models, showcasing significant advancements in the field of fake news detection on social networks.</p>