Placing user-generated content on the map with confidence

We describe a method that predicts the location of user-generated content using textual features alone. Unlike previous methods for geotagging text documents, our proposed method is not sensitive to how we discretize space. We also discover that spatial resolution has an impact on the prediction accuracy, which allows us to trade-off the spatial resolution of the predicted location against our confidence about its accuracy. Our method can be used to estimate the error in document's predicted location, enabling us to filter out poor quality predictions. We evaluate the proposed method extensively on user-generated content collected from two different social media sites, Flickr and Twitter. Our evaluation examines its performance on the geotagging task and with respect to different parameters. We achieve state-of-the-art results for all three tasks: location prediction, error estimation and result ranking and also provide a theoretical explanation of the effect of spatial resolution factor on geotagging accuracy. Our findings provide valuable insights into the design of geotagging systems and their quality control.