Strategies for combining Twitter users geo-location methods

Twitter has become a major player in the social media scene with over half billion users and over 500 million tweets published daily. With this abundant data, researchers saw the opportunity to explore this data for monitoring events and tracking epidemics. In this type of application, knowing the location of the user is essential. However, most of the information about location self-reported by users is difficult to process, and barely 1% of all published tweets are geolocated. Hence, user location inference is often performed by analyzing public available information from the user profile and his tweets. In this work, we evaluate and compare 16 approaches for user location inference based on different information sources that include interaction networks and text from tweets. We show that methods working with the user friendship network obtain higher values of accuracy and recall when compared to the other methods. From these results, we verify the agreement of pairs of methods regarding the predicted location and the users they cover. We find out that most methods disagree in their inferences while covering different sets of users. These results open up an opportunity to combine different methods in order to improve location accuracy and user recall. We propose four methods for combining the outputs of the evaluated methods. Two of them, one based on a weighting vote scheme (GAVe) and another based on a meta decision tree cover at least 98% of the users in the dataset, while location 75% of them within a distance of 100 km from their real location.