Long- and Short-term Preference Learning with Enhanced Spatial Transformer for Next POI Recommendation

Abstract

With the widespread adoption of mobile devices and Internet technologies, location-based social networks (LBSNs) provide a multitude of services to people. Next point of interest (POI) recommendation has become an important problem. The purpose of this task is to discover the location history activities from the user's preferences and recommend the next POI. The researchers utilized RNN model or attention mechanism to integrate long- and short-term interests and achieve success. However, existing works manually designed feature interaction to fuse different preference, or shallowly mined spatio-temporal information. To address the limitations, we propose an transformer-based model named Long- and Short-term Preference Learning with Enhanced Spatial Transformer(LSEST). Our model adopts a unified model to simultaneously model long-term and short-term preferences, so that the two user preferences can be interacted deeply to represent a comprehensive user preference. In addition, our model utilizes two transformer encoders to capture the temporal and spatial dependencies, respectively, and enhances the spatio-temporal consistency greatly. Extensive experiment on two real-world check-in datasets shows the superiority of our model compared to the state-of-the-art methods.