Prediction of Venues in Foursquare Using Flipped Topic Models

Foursquare is a highly popular location-based social platform, where users indicate their presence at venues via check-ins and/or provide venue-related tips. On Foursquare, we explore Latent Dirichlet Allocation (LDA) topic models for venue prediction: predict venues that a user is likely to visit, given his history of other visited venues. However we depart from prior works which regard the users as documents and their visited venues as terms. Instead we ‘flip’ LDA models such that we regard venues as documents that attract users, which are now the terms. Flipping is simple and requires no changes to the LDA mechanism. Yet it improves prediction accuracy significantly as shown in our experiments. Furthermore, flipped models are superior when we model tips and check-ins as separate modes. This enables us to use tips to improve prediction accuracy, which is previously unexplored. Lastly, we observed the largest accuracy improvement for venues with fewer visitors, implying that the flipped models cope with sparse venue data more effectively.