Predict User Mobility Pattern Based On Multivariate Transition Tensor

In the era of mobile Internet, the value of connecting people and services is becoming more and more important. Predicting the future mobile behavior based on the user's GPS trajectory data and then providing users with personalized service, has become an increasingly important research area.In this paper, we propose a multivariate transition tensor model, the model incorporated the user's time and location context and the interaction among the users compared with prior method which only considering user's location context. Based on the time-location sequence obtained by time discretization, the time-location two-variate transition tensor was established. The interinfluence between two users is quantified by correlation coefficient, and then the three-variate transition tensor is established based on the user's weighted influence model. The next-step distribution tensor is obtained by the three-mode multiplication of the current distribution tensor and the three-variate transition tensor. we can make short-term forecasting based on the distribution tensor which contains all users' spatial-temporal distribution. The stable distribution tensor is obtained by tensor contraction based power method and then we can make long-term forecasting.Multivariate transition tensor model takes into account the user, time and location factors and Incorporating multi-user spatial-temporal information and the mutual influence. The experiment based on a real user trajectory dataset is conducted and the results show that our model has better prediction accuracy than location or time transition model.