Inference models for opinion formation in social media
Il 3 febbraio 2020 alle 14.30
Sala Consiglio (ottavo piano) del Dipartimento di informatica dell'Università di Milano
Speaker: Corrado Monti, Postdoc Researcher, ISI Foundation.
Responsabile: Sebastiiano Vigna
Social media have significantly transformed how the general public consumes information. Despite the increasing attention received by the research community, modelling how opinions interact with social media remains an important open question. I am going to show how the design of such models can be tackled quantitatively thanks to maximum likelihood estimation on probabilistic graphical models.
The first question is about the space of opinions. We designed a model for how controversial content propagate on a social network, in which opinions live in a multidimensional, topic-based, ideological space. With a gradient-descent procedure, we can fit this model to data, turning a theoretical model for controversial content spreading into an algorithm for multi-dimensional opinion mining. In this framework we can infer people positions on different axes: economy, foreign policy, minorities, and study how different leanings relate to each other.
The second question is about how opinions evolve in time. Agent-based models are traditionally used for this aim, since they are easy to interpret, and encode sociological theories as causal mechanisms for opinion formation. However, these models do not exploit the widespread availability of social traces online. We show how by casting a traditional ABM as a problem of assigning signs on the edges of a temporal graph, we can perform inference and recover a likely latent opinion trajectory. This framework can then be used to analyze the evolution in time of internet communities, quantifying phenomena such as polarization and normalization.