Konstantinos Mylonas, "Recommendation driven opinion de-polarization in social networks", Diploma Work, School of Electrical and Computer Engineering, Technical University of Crete, Chania, Greece, 2024
https://doi.org/10.26233/heallink.tuc.101259
The Internet has acquired its role in the everyday life of people. Social media are now the ground for political debate and exchange of opinions. Initially, we might think that this over exposure to information would lead to open-minded, inclusive societies with less polarized members. However, there is a significant amount of work that suggests the opposite. People tend to organise into groups and communities that share common beliefs and interact with each other. Such communities are known as echo chambers. This phenomenon is known as homophily and has been studied for years by sociologists. Inside an echo chamber environment, the pre-existing beliefs of individuals are reinforced and the overall polarisation in the social network increases, which have negative impact to our society. In order to de-polarise the network, one can try to convince a small set of network members (users) to adopt more moderate positions around a topic. However, choosing a proper set of users is not an easy task, especially in the modern social networks that consist of billions of users. Should we focus on users with extreme opinions, but few connections? Or should we focus on less extreme, but famous, users, hoping that their neutral views will eventually reach and affect a larger fraction of the overall network? In this diploma thesis, we propose an efficient algorithm to choose a set of users from a network, such that when their opinions are moderated the overall polarisation of the network is reduced significantly, based on the famous Friedkin and Johnsen opinion formation model. We use Graph Neural Networks - specifically a Graph Convolutional Network - in order to create an algorithm able to work with large graphs. We compare our algorithm with a greedy algorithm, named GreedyExt, which has been proposed in the past. Our results show that our algorithm is much faster than GreedyExt and achieves similar performance in terms of depolarisation. We evaluate our algorithm in both synthetic and real graphs with ground truth communities.