Integração de informações de usuários para inferência de localização geográfica no twitter

Abstract

Knowing the location of a social-network user is essential to many applications that rely on the informations extracted from those networks, like detecting events and tracking epidemics. In this work, we investigate the task of inferring the location of Twitter users by recognizing that it may be encoded not only in their GPS-tagged messages, but also in their messages, profile information and the network of users they interact with. While previous works exploit the possibility of using those different sources of information, there is not a de facto technique that is a consensus as the best, but a range of many techniques with specific pros and cons. The sparsity of users in a geographic area may be a challenge to many inference methods due to the lack of enough data to characterize it. Therefore we initially discuss how to overcome the sparsity of users in many geographic regions. Grouping and merging location is one of the main approaches to solve this challenge, although by doing it we are giving up the precision of the inferred location. We propose a metric to evaluate the quality of the generated clusters and present a new strategy to group cities. Previous work have used different performance metrics to evaluate their geo-inferecing methods, making it difficult to compare them directly. We use standardized metrics in the same dataset, comparing the methods that use the users' social network, their publications and the information provided voluntarily in their profiles. We evaluate different network-based methods and compare their performance in two different types of users network: the friendship and the mentions network. We make evident the differences and similarities in using those two networks in the geoinference task. After pointing the changes in the vocabulary use in the stream of publication as the time passes, we propose a new method of inferring the location of a user's location based not only in the text but also considering the publication of her messages. We achieve better precision than the non-temporal similar approaches from the literature. We also evaluate the performance of using of the profile fields provided by the users and the periodic update of the models created to infere the location of new users. Finally, we evaluate the benefits of combining the information outputted by each method to improve the final performance the the geo-inference task.