计算机工程与应用 ›› 2024, Vol. 60 ›› Issue (11): 75-83.DOI: 10.3778/j.issn.1002-8331.2303-0168

• 理论与研发 • 上一篇    下一篇

融合全局特征的时空网络兴趣点推荐算法

李鹏飞,贺洋,毋建宏   

  1. 1.西安邮电大学 经济管理学院,西安 710061
    2.西安邮电大学 现代邮政学院,西安 710061
  • 出版日期:2024-06-01 发布日期:2024-05-31

Spatio-Temporal Network Interest Point Recommendation Algorithm Fusing Global Features

LI Pengfei, HE Yang, WU Jianhong   

  1. 1.School of Economics and Management, Xi’an University of Posts and Telecommunications, Xi’an 710061, China
    2.Modern Postal College, Xi’an University of Posts and Telecommunications, Xi’an 710061, China
  • Online:2024-06-01 Published:2024-05-31

摘要: 随着基于位置社交网络的迅速发展,兴趣点序列推荐逐渐成为近年来研究热点之一。针对现有推荐方法忽略签到数据中的全局信息,未充分考虑序列签到数据之间的时空间隔问题,提出一种融合全局特征的时空网络兴趣点推荐算法。该方法利用关系图神经网络获取签到数据异构网络图的全局特征,将时空门控融入传统门控结构中,融合全局特征对用户移动行为进行建模,再引入自注意力机制学习用户偏好向量表示。在两个真实数据集上进行实验比较与分析,实验结果表明所提方法推荐性能优于同类算法,验证了算法的有效性。

关键词: 兴趣点推荐, 门控循环单元, 关联图神经网络, 自注意力机制

Abstract: Recommendation of point of interest (POI) is one of the most popular topics in location-based social network (LBSN). The existing recommendation methods do not fully consider the deep influence of the spatial and temporal intervals between sequences of check-in data on the recommended sequences. They ignore the global information in the check-in data and focus on the local preferences in the recent check-in sequence of a single user. To address these problems, this paper proposes a global feature fusion based spatiotemporal network (GSTN) interest point recommendation algorithm. The method uses graph neural networks to obtain global features of the heterogeneous network graph of check-in data, and incorporates spatiotemporal gating into the traditional gating structure, fuses global features to model users’ mobile behavior, and then introduces a self-attentive mechanism to learn user preference vector representation. Finally, the experiments are carried out on two real datasets. The experiments show that the proposesd approach outperforms similar algorithms in terms of recommendation performance and verifies the effectiveness of the algorithm.

Key words: point of interest (POI) recommendation, gated recurrent unit, relational graph convolutional networks, self-attentive mechanism