Next POI Recommendation Based on Spatio-Temporal Auto-Correlation Network and Meta-Learning

In recent years, next point of interest (POI) recommendation has become an important technology in the development of location-based services. Existing methods based on Recurrent Neural Networks (RNNs), which have advantages in modeling the sequential transitions of user behavior, have shown good performance in next POI recommendation. However, RNNs have limitations such as difficulty in effectively capturing the associations between visited points on long trajectories and a failure to fully capture the cross-modal knowledge between POIs and categories. To address these issues, this paper proposes a next POI recommendation model based on spatio-temporal auto-correlation network. The model first uses graph embedding to learn the global transition information between POIs and integrates it into a multi-modal self-attention network to capture the complex temporal dependencies of user visit sequences. It then represents the local information with geographical adjacency relationships through spatial attention to capture the complex spatial dependencies of user visit sequences. Additionally, it is well known that the next POI recommendation process often faces severe cold-start problems, so a next POI recommendation model based on meta-learning methods is also proposed. Experimental results of the proposed scheme show its efficiency compared to other existing well-known schemes.