Technologies of recommender systems are being increasingly adopted by Location Based Social Networks (LBSNs) with the purpose of recommending Pointsof-Interest (POIs) to their users, and different contextual characteristics have been incorporated to enhance this process. Among these characteristics, the time at which users express their preferences (typically, by checking-in to different POIs) and ask for recommendations, is frequently referred as a first-order feature in this process. However, even when its influence on improving the accuracy of recommendations has been empirically demonstrated, time is still mainly considered through a monogranular representation (one-hour or one-day blocks). In this article, we introduce a POI recommendation approach based on a multigranular characterization of time, composed of hour, day-of-the-week, and month. Based on this concept, we propose two representations of user check-ins: one that directly extends a monogranular proposal of time for POI recommendations, and other based on a statistical representation of check-in distributions in time. For both representations, corresponding algorithms to compute user similarity and preference prediction are introduced. The experimental evaluation shows promising results in terms of accuracy and scalability.

recommender systems, point-of-interest, time-aware recommendation, location-based social network