Elections in Twitter Era: Predicting Winning Party in US Elections 2020 Using Deep Learning

Abstract

Twitter has become one of the leading platforms for people to express their political opinions during elections. Understanding public sentiment by on-ground election polling is expensive and time-consuming. Recently, tweets have gained popularity for analyzing public sentiment toward political parties. However, a large number of tweets have to be analyzed to simulate real-world elections effectively. Most previous works on Twitter sentiment analysis have trained on small tweet datasets, and most have failed to consider negative sentiment tweets. Considering only the positive tweets does not produce the actual on-ground sentiment. In this study, we have analyzed a large corpus of positive, negative, and neutral sentiment tweets about political parties in the US elections 2020 and predicted the election outcome. We constructed three distinct classification models using Bidirectional LSTM, GRU, and Hybrid CNN-LSTM to classify tweets into positive, negative, or neutral sentiments. Our results show that Bidirectional LSTM achieved an accuracy of $$95.99\%$$ , outperforming other deep learning approaches and related works. A custom net-score metric is proposed that considers both positive and negative sentiments. The Democratic Party outperformed the Republican Party with a higher net score, indicating that the Democratic Party is predicted to win. VADER algorithm was used to find the winning margin, which considered polarity, compound sentiment score, and retweet count. We found that the Democratic Party would lead by a marginal gap of $$+6.250\%$$ , which is close to the actual election results.