ABSTRACT In the past decade, Point of Interest (POI) recommendation has become one of the most important tasks in location‐based social networks (LBSNs). It is worth noting that LBSNs provide rich heterogeneous relationship information between users and POIs, including POI‐category relationships, POI–POI relationships, and user–POI relationships. However, most existing methods often yield biased and insufficient representations due to their inability to capture heterogeneous relationships, thus performing poorly against data sparsity and popularity bias. To this end, we proposed a POI Recommendation Method based on Contrastive learning and Spatial aggregation (PRMCS). The key to PRMCS is its ability to decompose the sophisticated user behavior into two parts. One is a routine part characterized by central‐based graphs built from historical trajectories based on spatial aggregation. Another is a preference part that obtains the users' historical preferences from the user check‐in records. Finally, the above two parts were fused to obtain the POI recommendation list. Experiments on two real‐world datasets demonstrated that PRMCS outperformed state‐of‐the‐art methods.
Yu et al. (Wed,) studied this question.