Enhancing Social Network Analysis Through NLP, Knowledge Graphs, and Deep Learning Technologies

Abstract

The rapid evolution of social media and recommendation systems has transformed social interactions and information dissemination, presenting both opportunities and challenges for social analysis. This thesis provides comprehensive methods for analysing social media content, detecting echo chambers, mitigating misinformation, and maximizing influence in social networks.

First, the thesis addresses a fundamental task of social network analysis. Text summarization with knowledge graphs is crucial for condensing and contextualising vast unstructured data, enhancing trend detection and decision-making. Two novel text summarization models, KATSum and AaKOS, are proposed, demonstrating the effectiveness of integrating knowledge graphs. KATSum is a Knowledge-aware Abstractive Text Summarization model that enhances the standard Seq2Seq model, while AaKOS is an Aspect-adaptive Knowledge-based Opinion Summarization model that generates personalized, aspect-oriented summaries from product reviews.

Second, the thesis investigates user behaviour modelling and information diffusion in online social networks. A novel user behaviour model, utilising deep neural networks and knowledge graphs, presents users' personalised actions after receiving influence messages. The proposed model serves as the foundation for the rest of the chapters. Based on this, a novel influence maximisation algorithm is then proposed based on the user behaviour model, which aims to maximise the influence in online social networks by considering the information alteration during the diffusion process.

Third, the focus shifts to investigating solutions for tackling misinformation detection, especially given how information alteration through reframing can lead to the development of misinformation. By adopting frame theories from communication studies, we aim to identify misinformation portrayed from factual information but presented misleadingly. To address this challenge, we propose a deep-learning-based model called FrameTruth, which leverages large language models to extract the framing of information, incorporating this as a crucial feature in misinformation classification. This approach delves into the nuanced manipulation of narrative frames, an under-explored area within the AI community. By leveraging the power of pre-trained large language models and deep neural networks, FrameTruth detects misinformation originating from accurate facts portrayed under different frames, utilising the various impacts of framing theory elements.

Finally, the thesis tackles the challenge of identification of echo chambers in online social networks by applying deep learning methodologies to model user beliefs based on historical message content and behaviours. A novel content-based framework is proposed, capable of detecting potential echo chambers by creating user belief graphs. This framework also demonstrates the evolution of user belief over time using the user behaviour model.