Analyzing and Predicting the US Midterm Elections on Twitter with Recurrent Neural Networks

We propose a method and a system that aim to gauge local support for the two major US political parties in the 68 most competitive House of Representative districts during the mid-term elections. We analyze tweets explicitly posted from locations within each district. To distinguish between Republican and Democratic tweets, we adopt a RNN-LSTM binary classifier which reached validation accuracy of 85% over individual tweets, despite the highly implicit and short content shared on the social network. The method was able to predict the correct winner on 60% of the highly competitive (and thus extremely hard to predict also with traditional methods) districts. The lower result at district level is also an indicator of the population bias of the Twitter platform with respect to the actual voters. The classifier architecture, along with the other methods and tools we propose, is domain- and language- independent and may be applied to any highly polarizing topic with enough social media activity.