Polarization in Social Systems: Dynamics of Quotes and Retweets Networks

The objectives of the thesis is to discriminate the impact of active polarization in online social networks, i.e. the presence of antagonist communities, and to characterise its dynamics in time. We focus on the social network X/Twitter and, precisely, on the Italian political and journalistic debate in the years 2018 to 2022. The two main actions that users on X/Twitter can adopt to publicly express their consent or dissent are retweets and quotes: when retweeting a user shares a tweet on their profile, while when quoting the user shares a tweet and adds a public text on it. To construct our dataset, we select all the tweets made by a selected sample of accounts that we refer to as leaders. We also select all the retweets and quotes the leaders received from other accounts: we refer to these new accounts as users. In this way, the natural structure of the data is that of a time-varying bipartite network among leaders and users, both for retweets and for quotes, from 2018 to 2022. We analyse the dataset by projecting the bipartite network on the layer of leaders with different metrics, e.g. cosine similarity. For this purpose we only use retweets, that we assume to model agreement among users and leaders, and we can thus implement a community detection algorithm to find the communities in the network. In this way, the communities represent sets of leaders that are similar according to the retweets they receive in each time window. We construct the distributions of retweets and quotes of users in each community. In particular, by analysing the entropy of these distributions for both retweets and quotes, normalised with respect to the number of clusters found in that time window, we highlight the difference between the use of quotes and retweets among communities. To this scope, we also analyse users activities when retweeting and quoting in time, showing how retweets and quotes carry different information and thus can model different types of interactions. We also analyse the response times of users when quoting, in order to cut properly the dynamics in appropriate time windows. Finally, we show how a dynamics is present in our system and we construct the flux matrices among communities, both for retweets and for quotes. We use flux matrices in order to characterise the presence of active polarization and give some evidence on the antagonism between clusters. In particular, we conjecture that a high flux of quotes among two communities may be an indicator of antagonism among two communities. To sum up, in this work we: (i) analyse the difference the role of retweets and quotes in terms of the information carried and conveyed in the dynamics; (ii) discriminate active polarization, intended as the presence of segregated communities with conflictual interactions, from passive polarization, intended as non-interacting communities. To do so, we exploit the ambivalent role of quotes, that can be used as a form of reinforcement or as a form of antagonism between users and leaders in the dynamics; (iii) characterise the dynamics of the communities found as well as their possible antagonism through quotes. This thesis falls in the realm of computational social science. The tools involved are network theory, information theory, community detection and opinion dynamics.