Political Discussion and Leanings on Twitter: the 2016 Italian Constitutional Referendum.

The recent availability of large, high-resolution data sets of online human activity allowed for the study and characterization of the mechanisms shaping human interactions at an unprecedented level of accuracy. To this end, many efforts have been put forward to understand how people share and retrieve information when forging their opinion about a certain topic. Specifically, the detection of the political leaning of a person based on its online activity can support the forecasting of opinion trends in a given population. Here, we tackle this challenging task by combining complex networks theory and machine learning techniques. In particular, starting from a collection of more than 6 millions tweets, we characterize the structure and dynamics of the Italian online political debate about the constitutional referendum held in December 2016. We analyze the discussion pattern between different political communities and characterize the network of contacts therein. Moreover, we set up a procedure to infer the political leaning of Italian Twitter users, which allows us to accurately reconstruct the overall opinion trend given by official polls (Pearson's r=0.88) as well as to predict with good accuracy the final outcome of the referendum. Our study provides a large-scale examination of the Italian online political discussion through sentiment-analysis, thus setting a baseline for future studies on online political debate modeling.