GeoTeCS: Exploiting Geographical, Temporal, Categorical and Social Aspects for Personalized POI Recommendation (Invited Paper)

The maturity of the smartphone and the World Wide Web (WWW) technologies have driven many social network applications which have facilitated people to share text and multimedia contents. The social networks that facilitate users to share the check-in (location visit) information are known as the location-based social networks (LBSN)s and provide various information for a recommendation problem that span beyond the user-location ratings, and comments. For instance, the time of the check-in, the category of the POI, the distance of POI from the user's home, the user's friends' visit to that place, and so forth. It's worthwhile to explore and efficiently integrate such information for the desired purpose. A Point of Interest (POI) recommendation system uses a user's historical check-in information from LBSNs and recommends the list of places that are potential for future visits. Many of the existing POI recommendation systems have focused on either of the temporal (time of the check-in), the geographical/spatial (distance between check-in locations), or the social (friendship, and trust based) aspects. Incorporation of all the major aspects (the categorical, the geographical, the social, and the temporal) of check-ins into a single model is barely explored by other studies. In this paper, we propose a fused model termed GeoTeCS (Geographical Temporal Categorical and Social) for personalized location recommendation. GeoTeCS uses the matrix factorization technique and an extension of the Multi-center Gaussian Model (MGM) to model the users' historical check-in behavior by fusing the major check-in aspects. The contributions of this paper are: (i) it proposes a matrix factorization based location recommender that incorporates all the major aspects-the categorical, the geographical, the social, and the temporal aspects into a single model and (ii) it extensively evaluates the proposed model against two real-world datasets - the Gowalla, and the Weeplaces, to illustrate its effectiveness.