Machine learning for nowcasting elections in Latin America based on social media engagement and polls

Abstract

Contemporary social media (SM) represents a new communication paradigm and has impacted politics and electoral campaigns. The mobilization of the Arab Spring social movements was attributed to SM platforms, as well as successful electoral campaigns such as those of Obama and Trump in the U.S. (2008, 2012, and 2016), the Brexit campaign in 2016, and the Bolsonaro campaign for the Brazilian presidency in 2018. Within this new scenario, the advantages of collecting SM data over traditional polling methods include the huge volume of available data, the high speed, and low costs. Hence, researchers are endeavoring to discover how to use SM for nowcasting election results. However, despite the alleged success, the most-common approach, based on counting the volume of mentions on Twitter and conducting a sentiment analysis, has been frequently criticized and challenged. On the other hand, recent approaches based on other SM platforms and on the advances of machine learning (ML) may be promising alternatives. In this context, this thesis aims to advance the state-of-the-art on predicting elections based on SM data. It proposes a new set of SM performance metrics to be input features for the ML techniques by changing the focus onto the number of people paying attention to the candidates. The defined metrics may be used not only with the most commonly-used current SM platforms (i.e., Facebook, Instagram, and Twitter) but even with future platforms which have not yet gained popularity. In addition, this thesis defines SoMEN, the Social Media framework for Election Nowcasting, a framework composed of a process and model for nowcasting election results based on the SM performance features and using ML approaches. It proposes well-defined steps, ranging from election understanding to prediction evaluation, and an ML model for predicting the final election results based on an ensemble of artificial neural networks (ANN) trained with SM metrics as features and offline polls as labeled data. It also defines SoMEN-DC, an execution strategy for SoMEN that enables continuous prediction during the campaign (DC). The proposed metrics and framework were applied on the most recent main presidential elections in Latin America: Argentina (2019), Brazil (2018), Colombia (2018), and Mexico (2018). More than 65,000 posts were collected from the SM profiles on Facebook, Twitter, and Instagram of the candidates, as well as data from 195 presidential polls. Results demonstrated that the defined metrics presented a high correlation with the final share of votes in all the studied countries. Moreover, it was also possible to achieve a high level of accuracy in predicting the final vote share of the candidates, with competitive or better results than traditional polls. Lastly, despite the difficulty in measuring the quality of predictions during the campaign, results are promising and also competitive to polls. The strategies put forward in this thesis have attempted to handle several among the current challenges in this research area and indicate a new manner on how to face the problems. Furthermore, they may be directly used for nowcasting future elections in similar scenarios.