Uma abordagem para seleção de tópicos relevantes em redes sociais online

Abstract

The increasing use of social networks has made the amount of information that is shared daily (personal opinions, news and events of the natural or social type) increasing significantly, making social networks a source of information about events. In Brazil, recent events such as Lava Jato Operation conducted by the Federal Police and the impeachment process of the Brazil President are examples of events that have had repercussions in the traditional medias. These facts can be used to characterize the relevant informations of an event. In social networks, the events generate discussions, sharings and news posts, where relevant information have more repercussions. However, processing a large mass of data to eliminate noise and recognize the information becomes costly. In this context, this work aims to show an approach to the characterization of relevant information by extracting topics in shared data on Twitter, where we evaluate the performance of three methods of machine learning (K-means, Latent Dirichlet Allocation (LDA) and Non-Negative Matrix Factorization (NMF)) used to extract topics on Lava Jato's databases and the impeachment process of the Brazilian president in two different preprocessing architectures (traditional and with entity recognition), demonstrating that it is possible to use the social networks as a data source to discover the relevant topics through the sensing of users that observe an event. In our experiments, we observed that the preprocessing techniques have a direct influence on the result of the topics extraction. In addition, we observed that the Silhouette technique helped us to find the best value of clusters for a given data sample. In the results the NMF presented the best performance in both databases, both in the topic extraction task and in the execution time