Location Representations for Accelerating the Training of Next POI Recommender System

With the widespread use of mobile devices, many next Point-Of-Interest (POI) recommender systems train users’ data on their devices to provide personalized recommendation services, which requires the algorithms to have an efficient convergence rate. Besides, POI location, an important feature in improving the POI recommendation performance, has been represented in various ways, such as encoding locations with distinctive IDs (e.g., different areas with different IDs). However, the positional relationship of a POI of interest and other POIs is ignored, which is crucial to predicting the next POI. In this study, we propose describing a POI’s location representation as the positional relationship with other POIs. Based on the idea that users prefer to visit closer regions, our approach transfers the distance interval of POIs to the similarity of the location representations, thus the possibility of a user visiting a POI can be revealed from the representation similarities. The proposed location representations that contain relative location information of all POIs can be applied to initialize POI recommendation models to accelerate model convergence. Experiments on our approach show that the proposed method on a sequential recommendation model improves Hit@10 by 4%, and the convergence rate (average regret before epoch@150) increases 4% with a 10% Hit@10 improvement.