Mining location information from users' spatio-temporal data

The availability of large localized datasets from popular social-networking sites such as Foursquare and Twitter has enabled the expansion of the field of human mobility analysis. Extracting location information from user trajectories helps to increase the understanding of human mobility and is important for improving public health, city planning, as well as crime prevention. In this work, we introduce a framework to automatically discover and classify locations based solely from geo-temporal data and without relying on contextual data, such as user explicit check-ins. We collected a dataset of 3 millions geo-tagged tweets of the New York City area to test our framework. By using a combination of clustering algorithms to accurately pinpoint location position (from a cloud of points representing tweets), and the popular time/days histograms as features to train the classifier, we were able to reach up to 53% average accuracy for the location type classification.