Exploiting user behavior learning for personalized trajectory recommendations

With increasing popularity of mobile devices and flourish of social networks, a large number of trajectory data is accumulated. Trajectory data contains a wealth of information, including spatiality, time series, and other external descriptive attributes (i.e., travelling mode, activities, etc.). Trajectory recommendation is especially important to users for finding the routes meeting the user’s travel needs quickly. Most existing trajectory recommendation works return the same route to different users given an origin and a destination. However, the users’ behavior preferences can be learned from users’ historical multi-attributes trajectories. In this paper, we propose two novel personalized trajectory recommendation methods, i.e., user behavior probability learning based on matrix decomposition and user behavior probability learning based on Kernel density estimation. We transform the route recommendation problem to a shortest path problem employing Bayesian probability model. Combining the user input (i.e., an origin and a destination), the trajectory query is performed on a behavior graph based on the learned behavior probability automatically. Finally, a series of experiments on two real datasets validate the effectiveness of our proposed methods.