Spatio-Temporal Self-Attention Network for Next POI Recommendation

Next Point-of-Interest (POI) recommendation, which aims to recommend next POIs that the user will likely visit in the near future, has become essential in Location-based Social Networks (LBSNs). Various Recurrent Neural Network (RNN) based sequential models have been proposed for next POI recommendation and achieved state-of-the-art performance, however RNN is difficult to parallelize which limits its efficiency. Recently, Self-Attention Network (SAN), which is purely based on the self-attention mechanism instead of recurrent modules, improves both performance and efficiency in various sequential tasks. However, none of the existing self-attention networks consider the spatio-temporal intervals between neighbor check-ins, which are essential for modeling user check-in behaviors in next POI recommendation. To this end, in this paper, we propose a new Spatio-Temporal Self-Attention Network (STSAN), which combines self-attention mechanisms with spatio-temporal patterns of users’ check-in history. Specifically, time-specific weight matrices and distance-specific weight matrices through a decay function are used to model the spatio-temporal influence of POI pairs. Moreover, we introduce a simple but effective way to dynamically measure the importances of spatial and temporal weights to capture users’ spatio-temporal preferences. Finally, we evaluate the proposed model using two real-world LBSN datasets, and the experimental results show that our model significantly outperforms the state-of-the-art approaches for next POI recommendation.