Hypergraph User Embeddings and Session Contrastive Learning for POI Recommendation

Abstract

Internet technologies have enabled location-based social networks (LBSNs) to provide users with a variety of services. In this context, next Point-of-Interest (POI) recommendation has become a key task. The goal of this task is to mine users’ travel behavior preferences based on their historical check-in session sequences and recommend the next POI. However, existing methods are insufficient in capturing the interactions between users and POIs, fail to fully mine personalized user preferences, and do not adequately reveal the complex, high-order collaborative signals between users, which makes it difficult to effectively alleviate the problem of data sparsity. Moreover, existing approaches also struggle to distinguish the different travel behavior patterns of users. To address these issues, we propose a novel method called Hypergraph User Embedding and Session Contrastive Learning (HUE-SCL) for next POI recommendation. We model users as hyperedges and the POIs they visit as nodes within these hyperedges, constructing a hypergraph that reveals the interaction relationships between users and POIs. Based on this hypergraph and POI embeddings, we perform personalized embedding of users to better capture their travel preferences and fully exploit the high-order collaborative signals between users, thus alleviating the data sparsity problem. Additionally, to better distinguish behavioral pattern differences between users, we introduce session enhancement and contrastive learning techniques to more accurately capture users’ travel preferences. Extensive experiments on two real-world datasets show that HUE-SCL outperforms existing state-of-the-art baseline methods.

Keywords

• Next POI recommendation
• hypergraph convolutional network
• contrastive learning
• transformer