politicanalytics: social media analytics for political data science
Luís Filipe Castanheira Gomes · Open Repository of the University of Porto (University of Porto) · 2014
In this project we studied the problem of predicting the results of political polls based on the combination of several aggregators of buzz and sentiment obtained from Twitter posts. We followed a machine learning approach, where the combination model was estimated from data. We used tweets from the portuguese tweetosphere since June 2011. This dataset contains tweets from 100 000 users who were classified as Portuguese. Futhermore, we had access to the polls results, since June 2011, of a private portuguese company that studies portuguese public opinion (Eurosondagens). We performed some experiments using two regression algorithms (Random Forests and Ordinary Least Squares). We compared the real poll values with the predicted values of our regression models. The lower absolute error we could obtain was 0.50 using only buzz aggregators. It means that our prediction model has a small predictive error. These results highlight the potential of using twitter data to complement or substitute traditional surveys.