Power of Predictive Analytics: Using Emotion Classification of Twitter Data for Predicting 2016 US Presidential Elections

Predictive analytics using the twitter feeds is becoming a popular field for research. A tweet holds a wealth of information on how an individual expresses and communicates their feelings and emotions within their social network. Large-scale collection, cleaning, and mining of tweets will not only help in capturing an individual’s emotion but also the emotions of a larger group. However, capturing a large volume of tweets and identifying the emotions expressed in it is a challenging task. Different classification algorithms employed in the past for classifying emotions have resulted in low-to-moderate accuracies thus making it difficult to precisely predict the outcome of an event. Secondly, the presence of diverse emotion annotated datasets, none of which are specific to a particular domain, has limited the potentiality of supervised algorithms for classification purposes. In this study, we demonstrate the potentiality of a lexicon-based classifier, NRC, which can mine emotions and sentiments in tweets. Using the NRC classifier, we initially determined the emotions and the sentiments within the tweets and used that to predict the swing direction of the 19 US states towards the candidates of the 2016 US presidential election. Comparing the predictions from the NRC against with the actual outcome of the election, we observed a ~90% accuracy, a performance superior to the mainstream pollsters indicating the potential emotion and sentiment-based classification holds in predicting the outcome of significant social and political events