As location-based applications rapidly gain popularity, a large volume of online contents with geo-tagged information are created daily. Check-ins, as a direct channel connecting the online and offline worlds, aid the development of many personalized and locational information services, such as personalized advertisement, local event promotion, and city management improvement.

Centered on Check-in behavior, this work aims to investigate the principles, methodologies and algorithms for knowledge discovery and predictive analysis in Location-based Social Networks (LBSN). Through collectively investigating geo-tagged information from multiple sources (social relationships, text information, biography information, etc.), several core tasks have been studied in this work, including modeling user behavior, geolocating social media users, Point Of Interest (POI) recommendation, etc.

Multiple techniques have been developed to fulfill the purpose of the above tasks. First of all, a goal-oriented co-clustering framework was proposed to incorporate customized information into the co-clustering process. Secondly, subspace learning technique was integrated into spectral co-clustering to incorporate multi-source information for venue recommendation. Thirdly, a generic geographical embedding model was proposed to learn the representation of objects in heterogeneous networks in LBSNs for geotagging social network users. Last but not the least, a deep content-aware POI recommendation model was proposed to structurally learn POI and user characteristics for intelligent POI recommendation.