Point-of-Interest (POI) recommendation is widely used in location-based services to help users discover potentially liking POIs. Existing methods mainly capture users’ preferences for POIs by learning the user check-in sequences or the user-POI bipartite graph, which either ignores the intrinsic spatio-temporal information of POIs or overlook the long-term spatio-temporal preferences of users, reducing recommendation performance. To address the above issues, in this paper we propose Spatio-Temporal Graph representation learning for POI Recommendation (STGRec). Specifically, we first construct a spatio-temporal graph to model the intrinsic relationships between POIs and learn the global embeddings of POIs from the spatio-temporal graph. Then, we learn the local embeddings of POIs by exploiting the spatio-temporal relationships between non-consecutive POIs in users’ check-in sequences, and we fuse the global and local embeddings of POIs to further capture users’ long-term spatio-temporal preferences, resulting in effective POI recommendation. Experiments on three real-world datasets show the proposed STGRec model outperforms the baseline models.