A Spatial-Temporal Self-Attention Network (STSAN) for Location Prediction

With the popularity of location-based social networks, location prediction has become an important task and has gained significant attention in recent years. However, how to use massive trajectory data and spatial-temporal context information effectively to mine the user’s mobility pattern and predict the users’ next location is still unresolved. In this paper, we propose a novel network named STSAN (spatial-temporal self-attention network), which can integrate spatial-temporal information with the self-attention for location prediction. In STSAN, we design a trajectory attention module to learn users’ dynamic trajectory representation, which includes three modules: location attention, which captures the location sequential transitions with self-attention; spatial attention, which captures user’s preference for geographic location; and temporal attention, which captures the user temporal activity preference. Finally, extensive experiments on four real-world check-ins datasets are designed to verify the effectiveness of our proposed method. Experimental results show that spatial-temporal information can effectively improve the performance of the model. Our method STSAN gains about 39.8% Acc@1 and 4.4% APR improvements against the strongest baseline on New York City dataset.