Feature-based POI grouping with transformer for next point of interest recommendation

With the increasing prevalence of location-based services, Point of Interest (POI) recommendation has become an active research topic. While Graph Neural Networks (GNNs) have been widely used in POI recommendation models, they suffer from computational efficiency limitations when the graph structure is large. In this paper, we propose a new next POI recommendation model, which is backboned by a lightweight, feature-based POI grouping (FPG) method and a Transformer network. A unique feature of the proposed model is it uses the FPG method, which divides POIs into multiple groups based on their geographical and popularity features and analyze the similarity among the users’ preferences on the groups. By using the FPG method rather than graph-based structures, the proposed model largely reduces the computational cost in making next POI recommendation. The POI embeddings generated by the FPG method are then fed into a Transformer to generate the recommendation result. We test the proposed model on three real-world datasets and conduct comprehensive comparison studies to validate the performance of the model. The experiment results show that the proposed model has superior computational efficiency while preserving sufficient next POI recommendation accuracy. Key findings and critical implications from the experiment result and the mechanistic design of the model are also discussed in detail.