Joint learning user's activities and profiles from GPS data

As the GPS-enabled mobile devices become extensively available, we are now given a chance to better understand human behaviors from a large amount of the GPS trajectories representing the mobile users' location histories. In this paper, we aim to establish a framework, which can jointly learn the user activities (what is the user doing) and profiles (what is the user's background, such as occupation, gender, age, etc.) from the GPS data. We will show that, learning user activities and learning user profiles can be beneficial to each other in nature, so we try to put them together and formulate a joint learning problem under a probabilistic collaborative filtering framework. In particular, for activity recognition, we manage to extract the location semantics from the raw GPS data and use it, together with the user profile, as the input; and we will output the corresponding activities of daily living. For user profile learning, we build a mobile social network among the users by modeling their similarities with the performed activities and known user backgrounds. Compared with the other work on solely learning user activities or profiles from GPS data, our approach is advantageous by exploiting the connections between the user activities and profiles for joint learning.