Uma Estratégia não Supervisionada para Previsão de Eventos usando Redes Sociais

Since the popularization of text-based social media, many studies have been conducted using data from these platforms to predict real-world events. The common approach is to consider people posting to a social media as sensors and their respective messages about an event as an indicator of its occurrence or intensity. In general, a sentiment analysis step is performed in order to assign messages to a set of predefined categories. In this case, manually labeled data is required to train a classifier. However, manually labeling messages can be costly, time consuming and is subject to human error. In addition, sometimes it is difficult to distinguish which predefined category is the most appropriate for an specific message. In this sense, we propose an unsupervised methodology to handle social media data for event prediction. We apply our methodology to dengue-related data from Brazil and show how an unsupervised approach can significantly improve the event prediction performance in the majority of the cases.