A Random Walk around the City: New Venue Recommendation in Location-Based Social Networks

The popularity of location-based social networks available on mobile devices means that large, rich datasets that contain a mixture of behavioral (users visiting venues), social (links between users), and spatial (distances between venues) information are available for mobile location recommendation systems. However, these datasets greatly differ from those used in other online recommender systems, where users explicitly rate items: it remains unclear as to how they capture user preferences as well as how they can be leveraged for accurate recommendation. This paper seeks to bridge this gap with a three-fold contribution. First, we examine how venue discovery behavior characterizes the large check-in datasets from two different location-based social services, Foursquare and Go Walla: by using large-scale datasets containing both user check-ins and social ties, our analysis reveals that, across 11 cities, between 60% and 80% of users' visits are in venues that were not visited in the previous 30 days. We then show that, by making constraining assumptions about user mobility, state-of-the-art filtering algorithms, including latent space models, do not produce high quality recommendations. Finally, we propose a new model based on personalized random walks over a user-place graph that, by seamlessly combining social network and venue visit frequency data, obtains between 5 and 18% improvement over other models. Our results pave the way to a new approach for place recommendation in location-based social systems.