EXTRACTING REAL SOCIAL INTERACTIONS FROM A DEBATE OF COVID-19 POLICIES ON TWITTER: THE CASE OF MEXICO

Abstract

Twitter is a popular social medium for sharing opinions and engaging in topical debates, yet presents a wide spread of misinformation, especially in political debates, from bots and adversarial attacks. The current state-of-the-art methods for detecting humans and bots in Twitter often lack generalizability beyond English. Here, a language-agnostic method to detect real users and their interactions by leveraging network topology from retweets is presented. To that end, the chosen topic is COVID-19 policies in Mexico, which has been considered by users as polemic. Two kinds of network are built: a directed network of retweets; and the co-event network, where a non-directed link between two users exists if they have retweeted the same post in a given time window (projection of a bipartite network). Then, single node properties of these networks, such as the clustering coefficient and the degree, are studied. Three kinds of users are observed: some with a high clustering coefficient but a very small degree, a second group with zero clustering coefficient and a variable degree, and a third group in which the clustering coefficient as a function of the degree decays as a power law. This third group represents ∼2% of the users and is characteristic of dynamical networks with feedback. The latter seems to represent strongly interacting followers/followed in a real social network as confirmed by an inspection of such nodes. A percolation analysis of the resulting co-retweet and co-hashtag network reveals the relevance of such weak links, typical of real social human networks. The presented methods are simple to implement in other social media platforms and can be used to mitigate misinformation and conflicts.