Point-of-interest (POI) recommendation in location-based social networks (LBSNs) predicts a user’s next visit using check-in records. As critical factors influencing user decision making, geographic and social contexts are incorporated to deliver higher-quality recommendation services. However, current methods suffer from two limitations: geographical modeling is confined to distance thresholds, ignoring long-range spatio-temporal transitions, and social graphs remain static, failing to capture dynamic behavioral similarities among unconnected users. To address these gaps, we propose GSTRDFA (Global Spatio-Temporal Relationship- and Dynamic Friendship-Aware), a model comprising three layers. First, we construct spatio-temporal KGs (STKGs) that encode four relationship types: global spatio-temporal and local geospatial POI–POI links, dynamic user–user friendships, and static social ties. Second, four dedicated encoders—STSEncoder (spatio-temporal state embedding), GeoEncoder (geographical convolution), DFEncoder (graph attention network), and SocEncoder (GraphSAGE)—propagate and aggregate user and POI embeddings along these STKG relations. Third, a GRU-based sequence predictor uses the fused embeddings to match candidate POIs to the user. Evaluations on Foursquare datasets (NYC, JK, CA) show that GSTRDFA outperforms existing methods, improving Acc@1/5/10 and MRR by 0.24–3.31%. Key contributions include (1) unifying spatial, temporal, and dynamic social signals via STKGs; (2) jointly modeling global spatio-temporal transitions and dynamic friendships; and (3) enabling balanced short-/long-range and short-/long-term transition prediction.

Point-of-interest (POI) recommendation in location-based social networks (LBSNs) predicts a user’s next visit using check-in records. As critical factors influencing user decision-making, geographic and social contexts are incorporated to deliver higher-quality recommendation services. While current methods suffer from two limitations: geographical modeling is confined by distance thresholds, ignoring long-range spatio-temporal transitions; and social graphs remain static, failing to capture dynamic behavioral similarities among unconnected users. To address these gaps, we propose GSTRDFA (Global Spatio-Temporal Relationship- and Dynamic Friendship-Aware), a novel model comprising three layers. First, we construct spatio-temporal KGs (STKGs) that encode four relationship types: global spatio-temporal and local geospatial POI-POI links, dynamic user-user friendships, and static social ties. Second, four dedicated encoders—STSEncoder (spatio-temporal state embedding), GeoEncoder (geographical convolution), DFEncoder (graph attention network), and SocEncoder (GraphSAGE)—propagate and aggregate user and POI embeddings along these STKG relations. Third, a GRU-based sequence predictor uses the fused embeddings to match candidate POIs to the user. Evaluations on Foursquare datasets (NYC, JK, CA) show that GSTRDFA outperforms existing methods, improving Acc@1/5/10 and MRR by 2.79-6.67%. Key contributions include: (1) unifying spatial, temporal, and dynamic social signals via STKGs; (2) jointly modeling global spatio-temporal transitions and dynamic friendships; and (3) enabling balanced short-/long-range and short-/long-term transition prediction.