Multi-view Contrastive Learning for Next POI Recommendation
摘要
Next POI recommendation aims to forecast their subsequent locations, which base been an important applications for LocationBased Services (LBS). Recently, there has been a growing interest in the investigation of POI recommendation using various deep learning techniques, and achieved great success. However, most of them only model the users preferences from a single view, such as the graph-structured features in the trajectory that indicate the preferences of individuals, or the dynamic change of personal preferences from trajectory sequences. Moreover, they ignore leverage the unique self-supervised signals to guide those two views of user preferences leaning perform better. To fill these critical voids, this paper proposes a Multi-View Contrastive Learning for next POI recommendation named MVCL-POI. Specifically, MVCL-POI first exploits dynamic-aware GCN from trajectory graph to automatically extract graph-structured features with considerations of rich node connections. After that, MVCL-POI utilizes a temporal-aware transformer to capture users’ dynamic interest from their historical trajectory sequences. Finally, a multi-view contrastive learning is designed to maximize the agreement between two views of users’ representations. Experiments results on two real-life LBSN datasets show that MVCL-POI yields significant performance advantages over the competitive next POI recommendation methods.
类型
出版物
Proceedings of the 2024 8th International Conference on Computing and
Data Analysis