Data-driven Analysis of Interactions and Popularity Increase in Online Social Networks

The rise of Online Social Networks in the last decade has changed society by irreversibly mingling the bounds between tangible reality and the online and, consequently, has shifted the paradigms of popularity relationships between famous entities and the public. It has been observed that Online Social Network (OSN) content popularity can be forecasted thanks to prediction algorithms applied to early metrics and that, more in general, predicting metrics can be both endogenous and exogenous to the OSN. This thesis introduces a novel approach to popularity forecasting based on historical information as well as attributes in control of the content creator instead of early popularity metrics and content quality attributes. We utilise data gathered between 2015 and 2021 about 1611 Instagram Italian influencer profiles from the Crowdtangle database, a public insights tool from Facebook that allows to follow, analyse, social media public content; such dataset comprises 2036966 posts, each characterised by the attributes generated by the users and metrics regarding its popularity. This dissertation proposes two algorithms, a Random Forest Regressor and a Recurrent Neural Network, implemented in several variations and some evaluation metrics with the goal of generating meaningful predictions about the number of reactions to a post without being subject to the extreme variance of metrics and the high number of outliers. The findings appear to indicate that a limit in the information contained in the data does not permit us to perform exact forecasts. Nonetheless, we are able to reach satisfactory results that usually predict future trends in the popularity of the influencer.