Characterizing geo-located tweets in brazilian megacities

This work presents a framework for collecting, processing and mining geo-located tweets in order to extract meaningful and actionable knowledge in the context of smart cities. We collected and characterized more than 9M tweets from the two biggest cities in Brazil, Rio de Janeiro and Sao Paulo. We performed topic modeling using the Latent Dirichlet Allocation model to produce an unsupervised distribution of semantic topics over the stream of geo-located tweets as well as a distribution of words over those topics. We manually labeled and aggregated similar topics obtaining a total of 29 different topics across both cities. Results showed similarities in the majority of topics for both cities, reflecting similar interests and concerns among the population of Rio de Janeiro and Sao Paulo. Nevertheless, some specific topics are more predominant in one of the cities.