Location-based social networks (LBSNs) have become more and more popular in the recent years. The typical LBSN platforms such as Foursquare, Facebook Local or Yelp allow the user to share their daily digital footprints in the form of check-ins with other people in different communities. The dynamic between users’ social context and their mobility plays an important role in LBSN, e.g. users potentially participate with their friends in the same activity. The social interaction also demonstrate in the form of multi-role context, as each user may experience different activities with each particular community. Existing representation learning for LBSNs analysis often fails to fully capture such complex social pattern. In this paper, we propose a representation learning model in which the multi-role social interaction can be captured simultaneously with the mobility information. More specifically, the model first applies a “persona” decomposition process, where each user node is splitted into several pseudo nodes presenting for his social roles. The process then learns multiple presentations for each persona that reflect the corresponding role by maximizing the collocation of nodes sampling on input user-user edges (friendships) and user-time-POI-semantic hyperedges (check-ins). We conduct experiments on 5 real-world datasets with 7 state-of-the-art baselines to demonstrate the robustness of our model on downstream tasks such as friendship suggestion and location prediction.