Developing a socio-computational approach to examine toxicity propagation and regulation in COVID-19 discourse on YouTube

Abstract As the novel coronavirus (COVID-19) continues to ravage the world at an unprecedented rate, formal recommendations from medical experts are becoming muffled by the avalanche of toxic content posted on social media platforms. This high level of toxic content prevents the dissemination of important and time-sensitive information and jeopardizes the sense of community that online social networks (OSNs) seek to cultivate. In this article, we present techniques to analyze toxic content and actors that propagated it on YouTube during the initial months after COVID-19 information was made public. Our dataset consists of 544 channels, 3,488 videos, 453,111 commenters, and 849,689 comments. We applied topic modeling based on Latent Dirichlet Allocation (LDA) to identify dominant topics and evolving trends within the comments on relevant videos. We conducted social network analysis (SNA) to detect influential commenters, and toxicity analysis to measure the health of the network. SNA allows us to identify the top toxic users in the network, which led to the creation of experiments simulating the impact of removal of these users on toxicity in the network. Through this work, we demonstrate not only how to identify toxic content related to COVID-19 on YouTube and the actors who propagated this toxicity, but also how social media companies and policy makers can use this work. This work is novel in that we devised a set of experiments in an attempt to show how if social media platforms eliminate certain toxic users, they can improve the overall health of the network by reducing the overall toxicity level.