Point of Interest Recommendation by Reality Mining Approach in Recommender Systems

Location-based social networks (LSBNs) are emerging services that have gained considerable popularity in recent years with the rapid advancement of mobile technology. LSBNs enable users to log their locations by recording entries, including various forms of contextual information (CI). Reality mining (RM) involves collecting and analyzing environmental and behavioral data from mobile devices to uncover predictable patterns, such as human mobility trends, which can enhance sequential next Point of Interest (POI) recommendations in recommender systems. Probabilistic models and sequence-based algorithms are among the most widely used approaches for learning mobility patterns in RM, although each presents its own challenges. In this study, for the first time, incorporate CI from LBSNs in a reality mining framework to predict users’ next POI within recommender systems. To this end, we propose a Contextual Extended Gated Recurrent Unit (CEGRU) architecture designed to separately investigate the impact of CI on location prediction. The CEGRU model extends the traditional GRU by introducing two distinct attention gates to better capture the impact of contextual variables on user movement behavior. Furthermore, this research introduces a novel experimental setup that evaluates model performance under two different dataset density conditions. This innovation enables the determination of the optimal dataset density for effectively assessing the proposed model. Comprehensive experiments were conducted on three large-scale real-world LBSN datasets, including Brightkite, Gowalla, and Foursquare. The results demonstrate that CEGRU outperforms competitive baseline methods on the Brightkite and Gowalla datasets in terms of Acc@10.