Predicting time-sensitive user locations from social media

Access to massive real-time user generated personal information from micro blogging services, such as Twitter and Facebook, has the potential to enable new location-based recommendation and advertising services. However, sparse user profile information and low adoption of per-message geo-coordinate information necessitates development of location detection techniques that exposes a user's location from message content. We propose and evaluate content-based machine learning techniques to a) identify tweets containing a user's location, and, b) categorize a user location into the author's present or future location. Such an approach is advantageous because it a) relies purely on message content, b) can be used to predict a user's future presence at a location, c) relates user locations to some context (activities, trip plans, etc.), and, d) can be used to profile users constantly evolving location. Our experimental evaluation shows that the proposed techniques can identify and categorize user locations from message content with high accuracy. We also extract the time entities associated with a user's future location to show when the user would be at that location. Finally we illustrate the location-based data analytics potential of these techniques on two real-world datasets.