<p>Next point-of-interest (POI) recommendation plays a crucial role in large-scale location-based services, aiming to predict the next location a user is likely to visit based on their historical trajectories. With the rapid growth of urban sensing data, modern POI recommendation systems must efficiently handle massive user trajectories and large POI graphs, which pose significant challenges in terms of computational efficiency and scalability. While existing methods leverage graph neural networks (GNNs) to model users’ historical trajectories, they often suffer from two key limitations. First, they struggle to effectively exploit the various latent factors underlying user mobility. Second, they neglect the global correlations among POIs. These limitations lead to an incomplete or biased understanding of user behavior, especially in large-scale scenarios. To overcome these limitations, we propose a novel dual-preference graph learning (DPGL) framework for next POI recommendation. Specifically, DPGL constructs two complementary graphs. The first is a local sequential graph that models users’ spatiotemporal transitions, while the second is a global geographical graph built based on POI distances. Subsequently, we capture the users’ sequential representations and the POIs’ geographical features based on the two graph structures, respectively. Furthermore, a self-attention mechanism and a long short-term memory (LSTM) network are introduced to the captured sequential features and geographical representations, allowing efficient modeling of complex dependencies in large-scale trajectory data and making the proposed framework amenable to parallel and high-performance computing architectures. Experimental results on three real-world datasets demonstrate that DPGL achieves superior performance compared with state-of-the-art baselines, offering an effective graph-based solution for large-scale next POI recommendation tasks.</p>