BREXIT: Predicting the Brexit UK Election Results by Constituency using Twitter Location based Sentiment and Machine Learning

Abstract

After parliament failed to approve his revised version of the ‘Withdrawal Agreement’, UK Prime Minister Boris Johnson called a snap general election in October 2019 to capitalise on his growing support to ‘Get Brexit Done’. Johnson’s belief was that he had enough support countrywide to gain a majority to push his Brexit mandate through parliament based on a parliamentary seat majority strategy. The increased availability of large-scale Twitter data provides rich information for the study of constituency dynamics. In Twitter, the location of tweets can be identified by the GPS and the location field. This provides a mechanism for location-based sentiment analysis which is the use of natural language processing or machine learning algorithms to extract, identify, or distinguish the sentiment content of a tweet (in our case), according to the location of origin of said tweet. This paper examines location-based Twitter sentiment for UK constituencies per country and aims to understand if location-based Twitter sentiment majorities per UK constituencies could determine the outcome of the UK Brexit election. Tweets are gathered from the whisperings of the UK Brexit election on September 4th 2019 until polling day, 12th December 2019. A Naive Bayes classification algorithm is applied to assess political public Twitter sentiment. We identify the sentiment of Twitter users per constituency per country towards the political parties’ mandate on Brexit and plot our findings for visualisation. We compare the grouping of location-based sentiment per constituency for each of the four UK countries to the final Brexit election first party results per constituency to determine the accuracy of location-based sentiment in determining the Brexit election result. Our results indicate that location-based sentiment had the single biggest effect on constituency result predictions in Northern Ireland and Scotland and a marginal effect on Wales base constituencies whilst there was no significant prediction accuracy to England’s constituencies. Decision tree, neural network, and Naïve Bayes machine learning algorithms are then created to forecast the election results per constituency using location-based sentiment and constituency-based data from the UK electorate at national level. The predictive accuracy of the machine learning models was compared comprehensively to a computed-baseline model. The comparison results show that the machine learning models outperformed the baseline model predicting Brexit Election constituency results at national level showing an accuracy rate of 97.87%, 95.74 and 93.62% respectively. The results indicate that location-based sentiment is a useful variable in predicting elections.