Scholarlink

<p>With the rapid development of location-based social networks, Point of Interest (POI) recommendation systems have emerged as a critical link connecting users to physical venues. However, current research predominantly simplifies POI recommendation into a sequential prediction task, thereby neglecting the collaborative signals inherent in user interactions. To address this limitation, we introduce a user-independent global trajectory flow graph that aggregates all user trajectories to capture common POI transition patterns, thus alleviating reliance on personalized features. Building upon this foundation, we adopt K-Nearest Neighbors (KNN) sparsification to construct modality-aware similarity graphs that preserve semantic POI correlations. Then, we design a graph-enhanced contrastive learning framework to refine multi-head Transformer-derived POI embeddings, achieving comprehensive utilization of collaborative signals for predictive modeling. Finally, we combine the traditional InfoNCE loss with graph representation learning objectives through weighted averaging, thereby enhancing feature expressiveness. Moreover, our method effectively mitigates cold-start problems in spatio-temporal recommendation systems. Experiments on two real-world datasets demonstrate significant superiority over state-of-the-art baselines.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Graph contrastive learning with global trajectory flow for next-POI recommendation

  • 期刊论文

  • 发表日期：

    2026年3月21日

查看全文
  • Hongwei Zhang,
  • Guolong Wang,
  • Keke Xu,
  • Kaijun Yang,
  • Wei Ni

摘要

With the rapid development of location-based social networks, Point of Interest (POI) recommendation systems have emerged as a critical link connecting users to physical venues. However, current research predominantly simplifies POI recommendation into a sequential prediction task, thereby neglecting the collaborative signals inherent in user interactions. To address this limitation, we introduce a user-independent global trajectory flow graph that aggregates all user trajectories to capture common POI transition patterns, thus alleviating reliance on personalized features. Building upon this foundation, we adopt K-Nearest Neighbors (KNN) sparsification to construct modality-aware similarity graphs that preserve semantic POI correlations. Then, we design a graph-enhanced contrastive learning framework to refine multi-head Transformer-derived POI embeddings, achieving comprehensive utilization of collaborative signals for predictive modeling. Finally, we combine the traditional InfoNCE loss with graph representation learning objectives through weighted averaging, thereby enhancing feature expressiveness. Moreover, our method effectively mitigates cold-start problems in spatio-temporal recommendation systems. Experiments on two real-world datasets demonstrate significant superiority over state-of-the-art baselines.