Building a Location-Based Set of Social Media Users

In many instances, one may want to gain situational awareness in an environment by monitoring the content of local social media users. Often the challenge is how to build a set of users from a target location. Here, we introduce a method for building such a set of users by using an expand–classify approach, which begins with a small set of seed users from the target location and then iteratively collects their neighbors and classifies their locations. We perform this classification using maximum likelihood estimation on a factor graph model that incorporates features of the user profiles and social network connections. We show that maximum likelihood estimation reduces to solving a minimum cut problem on an appropriately defined graph. We are able to obtain several thousand users within a few hours for many diverse locations using our approach. Using geolocated data, we find that our approach typically achieves good accuracy for population centers with fewer than 500,000 inhabitants but is less effective for larger cities. We also find that our approach is able to collect many more users with higher accuracy than existing search methods. Finally, we show that, by studying the content of location-specific users obtained with our approach, we can identify the onset of significant social unrest in locations such as the Philippines.