Meta-path interaction and decomposition for heterogeneous information networks-based POI recommendation

With the rapid development of mobile Internet and location-based services, personalized points of interest (POIs) recommendation has become one of the core tasks in location-based social networks. To address the limitations of existing recommender systems in modeling high-order contextual semantics and differentiating features across meta-paths, we propose MID-HINRec, a POI recommendation framework for heterogeneous information networks that leverages meta-path interaction and decomposition. The method integrates geographic, temporal, and semantic information, constructing multiple semantic paths to capture the complex user-POI associations. On this basis, non-negative matrix factorization and a dual-channel interaction fusion network are used to extract global semantic interaction features. We further model local structures by applying a heterogeneous graph attention network to heterogeneous subgraphs and a weighted graph convolutional network to homogeneous subgraphs. Furthermore, information entropy is introduced as prior attention weights to fuse multi-path local structural features. Finally, global and local features are jointly trained to predict user preferences. By integrating user and POI representations from multiple perspectives, the proposed model improves both recommendation accuracy and interpretability. Experiments on two real-world datasets (New York and Tokyo) show that MID-HINRec consistently outperforms the strongest baselines, achieving average relative gains of 3.31% in HR and 9.24% in NDCG.