A Dynamic-aware Heterogeneous Graph Neural Network for Next POI Recommendation

Next point-of-interest (POI) recommendation is of great importance for both location-based service providers and users. Current state-of-the-art methods view users and POIs as unified latent representations, and model users’ transition patterns from global and local views. However, most of them still have following limitations: 1) Ignoring user’s dynamic behavioral intention, which is significantly influenced by current temporal and spatial factors. 2) Insufficiently considering different activity connotations of POIs in various temporal contexts. To tackle these challenges, we propose a novel method Dynamic-aware Heterogeneous Graph Neural Network (DyHGN) for next POI recommendation, which jointly learns fine-grained representations from global and local views. In the global view, we first construct a series of dynamic-aware heterogeneous graphs, and design a fine-grained temporal enhanced graph neural network to learn users’ dynamic behavioral intentions and POIs’ dynamic activity connotations. In the local view, we propose a dynamic information aggregation module that employs a well-designed information enhancement layer to enhance robustness of the model. Furthermore, we improve the attention mechanism to learn important spatio-temporal factors in users’ behavior. Extensive experimental results on two real-world public datasets demonstrate the effectiveness of our proposed method.