Group-Aware Dynamic Graph Representation Learning for Next POI Recommendation

Abstract

The Next POI recommendation, which has attracted many attentions recently, is a complex study due to the sparsity of check-in data and numerous sequential patterns. The recent methods based on sequential modeling have shown promising applicability for this task. However, most of existing next POI recommendation researches only model users’ preferences based on their own sequences and ignore the influence of partners who visit POI with the target user and may change with time. Inspired by dynamic Graph neural networks, we propose a Group-aware Dynamic Graph Representation Learning (GDGRL) method for next POI recommendation. GDGRL connects different user sequences and specific partners via dynamic graph structure, which contains interactions between users and POIs as well as influence of partners. The users’ dynamic preferences are learned from group-aware dynamic graph and context-aware dynamic graph through dynamic graph neural networks. Finally, the next POI recommendation task is transformed into a link prediction between user node and POI node in the dynamic graph. Extensive experiments on two real-world datasets show that GDGRL outperforms several state-of-the-art approaches.