Characterizing community structures on social media over time: a graph learning approach

Abstract

In an age where information is more accessible than ever, it’s easy to assume that people are becoming more informed and open-minded. In spite of that, people are increasingly finding themselves in echo chambers, surrounded by like-minded individuals and exposed mainly to content that reinforces their existing beliefs. There are, however, social media users that break with that pattern by changing the group of users they interact with over time. In this study, we analyze the dynamics of interactions between users on Twitter and Reddit over extended periods, with the aim of identifying changes in community structures. We leverage the data available through these platforms’ APIs to construct user interaction graphs and use several methods to classify users into communities, including SBM, Infomap and Louvain, to classify users into communities. Additionally, we use NLP techniques such as Community Pooling, BERTopic and Perspective [8], as well as graph algorithms, to characterize different user profiles in online debates. Our research analyzes how social media communities and their users evolve over time, with implications for understanding online discourse and facilitating healthy interactions on these platforms. As a first approach, we analyzed three months of Donald Trump’s tweets, finding clear signs of polarization. Regarding the user flow between communities, we found that most of the users who changed communities twice went back to their original one (∼ 96%).