POI recommendation based on social trust in LBSN

Abstract

This thesis addresses the challenges of point-of-interest (POI) recommendation systems in location-based social networks (LBSNs), such as Yelp or Foursquare, with a focus on data sparsity and cold-start problems. To overcome these challenges, the thesis proposes several approaches based on the exploitation of implicit trust between users. Unlike declared friendship links (explicit trust), implicit trust is inferred from users’ behavior, particularly through their check-ins and ratings. Three main models were developed to integrate this trust into recommendation systems: (1) the HRCT model (Hybrid Rating Check-in Trust), which combines ratings and check-ins to build a denser trust matrix, there by reducing data sparsity and improving recommendation accuracy; (2) the PRCT model (Propagation of Rating/Checkin for implicit Trust), an extension of the HRCT model that applies a trust propagation mechanism within the social network, helping to mitigate cold-start issues; and (3) the ITCRC model (Implicit Trust based on Combining point of interest Ratings and user Check-ins), which incorporates trust directly into the POI prediction process. The experimental results, obtained from real-world datasets such as Yelp, showed that these models help to densify the similarity matrices and improve the accuracy of POI rating predictions based on user check-ins, while also reducing the impact of sparsity and the cold start problem. In particular, approaches that incorporate check-ins into the computation of the implicit trust matrix between users proved to be more effective than those based solely on ratings