Research On Location Prediction Based On Analysis Of User Behavior

In the past few years, the wireless communication networks have been rapidly developed and intelligent devices have been universally used. And with the improving of mobile position technology and the rising of location-based service (LBS), a new online social media, location-based social networks (LBSN), combing LBS and social network service (SNS),appears in our life. LBSNs accumulate a large number of user check-ins information for its function that allows users to share their locations anywhere and anytime. The trajectory data in LBSN implies user mobile behaviors and habits, and the social data implies social behaviors and interrelationship between users.The LBSN combines the users' online virtual identities with their offline real identities, and brings us the convenience of exploring human behavior in real life. Modeling users' behavior in LBSN can effectively insight into offline users' behavior motivations, understand the users'mobility patterns, and mine the universality and individuality of people's behaviors. Then based on these, we can predict future location for users.Over the past years, location prediction has already benefited many areas in our life, such as targeted advertising, urban planning, resource pre-allocating and so on. Therefore, research on human mobility and location prediction is essential for guiding various industries to deploy works in advance and promote business benefits. But the incompleteness and sparsity of human mobile data and the insufficient investigation of user behavior severely limit the accuracy of existing location prediction algorithms. Based on a comprehensive analysis of the LBSN users'mobility behavior, this thesis deeply studies the location prediction algorithm. The main work exhibits as follows:(1) Based on feature tendency, this thesis proposes a location prediction algorithm, Powerful Feature Tendency Based Selection (PFTS),which is universal for all datasets with diverse features. Furthermore, the proposed model can self-adaptively select the feature with the most powerful tendency by comparing variance of visiting probabilitydistribution. Based on the chosen features, the PFTS automatically employs corresponding method to predict locations for various scenes.Using two real-world LBSNs datasets, the experiment results validate that the PFTS significantly outperforms state-of-the-art approaches in terms of prediction accuracy. The results also suggest that the feature tendency captures users' mobile behavior for boosting prediction performances effectively.(2) Based on the community detection in complex networks, this thesis proposes two location prediction algorithms. It is proved that user clustering based on behavior similarity can effectively capture the community structure in the network. Thus, the trajectory patterns are modelled to determine the similarity of user geographic behaviors, and the social patterns are modelled to determine the similarity of social behaviors.Secondly, a T+S model and a TS model are clustered and merged users'data respectively. Finally based on community, we predict locations for users. The results show that user clustering can significantly improve the prediction accuracy.