Hybrid Model Recommended for the Next POI

As science and technology progress rapidly and smart devices become increasingly prevalent, the importance of recommendation systems across various fields has become increasingly prominent, and the demand for next POI (Point of Interest) recommendations is also rising. This paper presents a hybrid model for next POI recommendation, which learns multiple user-region matrices via multi-task learning, followed by Hierarchical Beam Search (HBS), to predict the next POI. In this process, the attention mechanism and layer normalization are integrated to effectively capture global POI relationships. The POIs in the spatiotemporal graph are linked by region and timeslot, tackling the data sparsity issue. The analysis suggests that the approach exceeds the performance of existing recommendation methods.