Inferring Twitter user locations with 10 km accuracy

Geographic locations of users form an important axis in public polls and localized advertising, but are not available by default. The number of users who make their locations public or use GPS tagging is relatively small, compared to the huge number of users in online social networking services and social media platforms. In this work we propose a new framework to infer a user's main location of activities in Twitter using their textual contents. Our approach is based on a probabilistic generative model that filters local words, employs data binning for scalability, and applies a map projection technique for performance. For Korean Twitter users, we report that 60% of users are identified within 10 km of their locations, a significant improvement over existing approaches.