Research Of Personalized Recommendations On Location-based Social Networks

With the development of the Internet and the popularity of mobile devices, location-based social network services(LBSNs for short), such as Gowalla and Foursquare have greatly facilitated our life. An LBSN allows users to establish social links, check in points-of-interest(POIs), and rate their visiting POIs. These historical check-in data bring new opportunities to mine user preferences for personalized POI recommendations. How to make high quality recommendations on such a platform is becoming a very valuable problem.As geographical information plays an important role in location-based recommendations, we propose three models, namely, ILARS-MF model, ULARS-MF model and UILARS-MF model to explore the geographical influence in recommender systems via matrix factorization. Here, ILARS-MF corresponds to non-spatial ratings for spatial items; ULARS-MF corresponds to spatial ratings for non-spatial items and UILARS-MF corresponds to spatial ratings for spatial items. Current studies used different methods to make recommendations according to these three kinds of ratings. Due to the fact that all three types of ratings are significant and matrix factorization is an effective method in fusing different factors for recommendations, we propose a unified framework for location-aware recommender systems with the consideration of geographical influence using the matrix factorization method.In fact, users always check in few venues, the sparse data in LBSN provide us with a big challenge to make high quality recommendations. To solve such problems, in addition to geographical influence, we also incorporate social influence and categorical influence into our recommendations. Through analyzing on real-world data set Foursquare, we find all these three factors follow power law distributions. As the power law distribution is an important and widely used distribution, we model social, categorical and geographical influence as power law distributions, respectively. Then we combine these three factors in a unified framework and use gradient descent methods to explore the weight parameters of our recommendation model.We compare our methods with other recommender methods based on real-world data sets and the experimental results show that our methods have a better result.