Leveraging knowledge graph and mining user multiple preferences for next point-of-interest recommendation

Next point-of-interest (POI) recommendation, which leverages user check-ins to predict their most probable next destination, becomes the core function of location-based services. Most previous approaches derive static features from users’ past and current check-ins to capture both long- and short-term POI preferences. However, these methods fail to fully mine dynamic contextual interactions in user sequences, limiting their ability to effectively model user behavior patterns. Moreover, they overlook users’ diverse preferences, which reflect interests at different levels and influence recommendation performance. To address these two challenges, this paper proposes an algorithm that leverages a location transition knowledge graph and extracts user multiple preferences for next POI recommendation (KGMP). First, a location transition knowledge graph is constructed to model users’ static check-ins and dynamic transition relationships between POIs. Then, user intrinsic and contextual POI preferences are extracted from the knowledge graph using a GRU model and a Transformer module, respectively. Additionally, users’ preferences for POI categories are obtained from their category sequences using an attention mechanism. Finally, POI recommendation lists for users’ next visits are generated by linearly integrating user multiple preferences. Extensive experiments on four real-world datasets demonstrate that KGMP achieves average 23.49% improvement in terms of NDCG@10 over the best baseline STKGRec.