Scholarlink

<p>With advancements in high-precision positioning technology and the availability of commercial map services, LBSNs can now acquire both geographic coordinates and precise category information of locations. However, most existing models for user location prediction suffer from two significant limitations: neglecting category information and overlooking the hierarchical relationship between specific locations and their corresponding categories. To overcome these challenges, we propose the hierarchy aware-based multi-task learning (HAMTL) framework, which jointly predicts the next location and its category. HAMTL extracts hierarchy information by constructing a hierarchy tree, which is then fused with other feature information before encoding. Finally, a hierarchical decoder is designed for prediction. Extensive experiments conducted on two real-world datasets demonstrate that HAMTL outperforms seven baselines across all evaluation metrics, achieving a maximum <i>Acc</i>@1 improvement of over 18% compared with the state-of-the-art method. In addition, we show that the hierarchical decoder provides advantages over HAMTL on the basis of the effective modeling of hierarchy information.</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

Hierarchy aware-based multi-task learning for user location prediction

  • 期刊论文

  • 发表日期：

    2025年7月29日

查看全文
  • Yahui Wang,
  • Hongchang Chen,
  • Shuxin Liu,
  • Junjie Zhang,
  • Lan Wu,
  • Xiaoyan Cui,
  • Yuxiang Hu

摘要

With advancements in high-precision positioning technology and the availability of commercial map services, LBSNs can now acquire both geographic coordinates and precise category information of locations. However, most existing models for user location prediction suffer from two significant limitations: neglecting category information and overlooking the hierarchical relationship between specific locations and their corresponding categories. To overcome these challenges, we propose the hierarchy aware-based multi-task learning (HAMTL) framework, which jointly predicts the next location and its category. HAMTL extracts hierarchy information by constructing a hierarchy tree, which is then fused with other feature information before encoding. Finally, a hierarchical decoder is designed for prediction. Extensive experiments conducted on two real-world datasets demonstrate that HAMTL outperforms seven baselines across all evaluation metrics, achieving a maximum Acc@1 improvement of over 18% compared with the state-of-the-art method. In addition, we show that the hierarchical decoder provides advantages over HAMTL on the basis of the effective modeling of hierarchy information.