Collective attention in online social networks

Abstract

Social media is an ever-present tool in modern society, and its widespread usage positions it as a valuable source of insights into society at large. The study of collective attention in particular is one application that benefits from the scale of social media data. In this thesis we will investigate how collective attention manifests on social media and how it can be understood. We approach this challenge from several perspectives across network and data science. We first focus on a period of increased media attention to climate change to see how robust the previously observed polarised structures are under a collective attention event. Our experiments will show that while the level of engagement with the climate change debate increases, there is little disruption to the existing polarised structure in the communication network. Understanding the climate media debate requires addressing a methodological concern about the most effective method for weighting bipartite network projections with respect to the accuracy of community detection. We test seven weighting schemes on constructed networks with known community structure and then use the preferred methodology we identify to study collective attention in the climate change debate on Twitter. Following on from this, we will investigate how collective attention changes over the course of a single event over a longer period, namely the COVID-19 pandemic. We measure how the disruption to in-person social interactions as a consequence of attempts to limit the spread of COVID-19 in England and Wales have affected social interaction patterns as they appear on Twitter. Using a dataset of tweets with location tags, we will see how the spatial attention to locations and collective attention to discussion topics are affected by social distancing and population movement restrictions in different stages of the pandemic. Finally we present a new analysis framework for collective attention events that allows direct comparisons across different time and volume scales, such as those seen in the climate change and COVID-19 experiments. We demonstrate that this approach performs better than traditional approaches that rely on binning the timeseries at certain resolutions and comment on the mechanistic properties highlighted by our new methodology.