Detecting Influence in Online Social Networks using Conversational Features.

Abstract

Social media is one of the most common forms of communication today. Unlike traditional face-to-face dialogue, a single user on social media can impact the lives of millions of people, either directly or indirectly. In today’s increasingly connected world, social network analysis research has been critical to understanding new and emerging social phenomena, including the reach of highly-influential users. However, these methods fail to identify the initial causes of influence by only assessing its after-effects via changes to network topology. Content, the originator of influence in many cases, has until recently been considered an unlikely replacement for social network analysis. Prior to the introduction of attention mechanisms and the transformer architecture, this had been attributed to the unbalanced ratio of its high resource requirements and comparably low predictive capabilities. However, despite the field’s recent growth, there is limited research evaluating: whether these capabilities transfer to detecting interpersonal influence, the relationship between content-informed models and traditional topological features, and the corresponding traits most associated with influence via a content-based approach. This thesis combines these two fields to assess whether a content-led method can offer a suitable substitute for existing topological techniques. To evaluate this, I adopt several state-of-the-art large language models, a novel fine-tuned feature extraction framework, two newly collected social media datasets and a large array of topological metrics and conversational features for comparison. Broadly, this thesis shows that content-led approaches can offer a comparatively accurate assessment of influence under certain conditions, and in doing so they also provide researchers with greater insight into influence dependencies on social media.

Date of Award	17 Jun 2025
Original language	English
Awarding Institution	University of Bristol
Supervisor	Matthew Edwards (Supervisor) & Ryan McConville (Supervisor)