Capturing Geographical Influence in POI Recommendations

Point-of-Interest (POI) recommendation is a significant service for location-based social networks (LBSNs). It recommends new places such as clubs, restaurants, and coffee bars to users. Whether recommended locations meet users’ interests depends on three factors: user preference, social influence, and geographical influence. Hence extracting the information from users’ check-in records is the key to POI recommendation in LBSNs. Capturing user preference and social influence is relatively easy since it is analogical to the methods in a movie recommender system. However, it is a new topic to capture geographical influence. Previous studies indicate that check-in locations disperse around several centers and we are able to employ Gaussian distribution based models to approximate users’ check-in behaviors. Yet centers discovering methods are dissatisfactory. In this paper, we propose two models—Gaussian mixture model (GMM) and genetic algorithm based Gaussian mixture model (GA-GMM) to capture geographical influence. More specifically, we exploit GMM to automatically learn users’ activity centers; further we utilize GA-GMM to improve GMM by eliminating outliers. Experimental results on a real-world LBSN dataset show that GMM beats several popular geographical capturing models in terms of POI recommendation, while GA-GMM excludes the effect of outliers and enhances