Capturing Fine and Coarse Grained User Preferences with Dual-Transformer for Next POI Recommendation

Abstract

The next Point of Interest (POI) recommendation is of great significance to provide personalized location suggestions for users to visit next. Most existing researches fail to explore user coarse-grained category movement preferences embedded in the check-ins, and ignore more factors that affect POI relationship construction, so as not to effectively deal with the issue of data sparsity, and limit the model’s generalization ability. To address the above challenges, a next POI recommendation method is developed, which captures fine and coarse grained user preferences with Dual-Transformer (FCDT). Specifically, we construct a POI relationship directed graph based on all user check-in trajectories and Gated Graph Neural Network(GGNN) is employed to generate a more expressive POI embedding, then it is input into the Transformer model to learn the user’s fine-grained POI preferences in POI sequences. Next the user’s coarse-grained preference on the category level is learned based on the Transformer framework. Finally, the adaptive weighted-sum method is used to integrate the user’s fine and coarse-grained preference to calculate the user’s preference probability of POI. A number of experiments conducted on three real world datasets demonstrate the superiority of the proposed FCDT algorithm, it has better performance compared with other six algorithms in terms of accuracy and mean reciprocal rank.