Mobility prediction using pattern network

Abstract

Nowadays, location-based services are ubiquitous in our daily life. Many mobile applications recommend nearby restaurants, transportations or new places based on user's profile. However, considering the privacy protection, customers are less willing to provide detailed information, such as age, status, habits etc.. Only moving trajectories may be accessible by telecom service providers. Mining in trajectories is one of strategies to understand customer's behaviors. Therefore, how to discover user's mobility patterns as well as making accurate mobility prediction become two critical issues for location based services. Moreover, trajectories that are updated frequently from massive people behave as streaming data, which require pattern mining and prediction algorithms to be efficient as well. In this thesis, we introduce our methods to conquer the above challenges. Our first work is to find atomic mobility patterns from trajectories. Since moving trajectory usually consists of many tandem repeats, the proposed pattern mining algorithm is able to perform repeating sub-sequence mining and tandem structure detection concurrently. With a pipeline framework, we can discover various patterns in an online manner. Next, we transform a location sequence to a novel pattern-based network by connecting all discovered patterns. The pattern network models user's historical movements from location level to pattern level, which not only provides a graph presentation for investigating user's mobility, but also serves as a mobility model for better prediction. Our pattern network model is trained by three steps including prediction, verification and weight propagation. Through online tunning the parameters of pattern network, user's next location can be predicted in real time. Finally, we focus on the mobility prediction of unusual behavior. The motivation is that many locations in our daily life are visited infrequently or only once. Usually, these locations are hard to be predicted successfully by traditional methods. We introduce the concept of mobility change, called Point of Change (POC), to describe people's new and unusual mobility behaviors. Our pattern network model is extended to include spatial-temporal information for learning and predicting possible POCs in a user's trajectory. In general, our experiments show that the pattern network model outperformed other Markov models on location prediction and unusual mobility prediction. Moreover, people's mobility behaviors can be further analyzed according to the structure of pattern network.