Journal of Frontiers of Computer Science and Technology ›› 2023, Vol. 17 ›› Issue (7): 1680-1689.DOI: 10.3778/j.issn.1673-9418.2202014

• Artificial Intelligence·Pattern Recognition • Previous Articles     Next Articles

Method of Heterogeneous Graph Representation Learning Using Novel Meta-Path

ZHANG Chengdong, WANG Shaoqing, LIU Yufang, ZHENG Shun, SUN Fuzhen   

  1. School of Computer Science and Technology, Shandong University of Technology, Zibo, Shandong 255000, China
  • Online:2023-07-01 Published:2023-07-01

采用新型元路径的异构图表示学习方法

张程东,王绍卿,刘玉芳,郑顺,孙福振   

  1. 山东理工大学 计算机科学与技术学院,山东 淄博 255000

Abstract: Graph neural network has become a main method in the field of recommendation system. Many studies integrate meta-path into heterogeneous graph neural network. However, most of the definitions of meta-paths only consider whether there is edge connection between nodes, but the same node may be connected by multiple different types of edges in the heterogeneous graph, such as users’ browsing, adding to the shopping cart, purchasing and other interactive behaviors. Then, instantiating according to the traditional definition method of meta-path can result in inaccurate embedding of nodes because the types of edge are ignored. To solve this problem, this paper proposes a method of integrating edge types into meta-paths in heterogeneous graphs, so that nodes can be trained indiv-idually in each scene. Then, the graph attention mechanism is used to aggregate the embedding of the same node in different scenes, and finally the embedding of the node is obtained. It is used to predict the behavioral relationship between users and non-interactive items, and then items are recommended to users. Experimental results show that the proposed algorithm improves the performance on three public datasets. In particular, the F1, ROC-AUC and PR-AUC scores are improved by 8.75%, 6.03% and 4.86% on Ali Tianchi Challenge dataset, respectively.

Key words: distributed representation, heterogeneous graph, meta-path, recommendation system

摘要: 图神经网络已经成为推荐系统领域的一种主要方法。很多研究把元路径融入到异构图神经网络中,但绝大多数元路径的定义方式只考虑节点之间是否存在连接。而在异构图中同一个节点可能被多条不同类型的边所连接,如用户对物品的浏览、加入购物车、购买等不同交互行为,按照传统的元路径定义方式进行实例化会因为忽略了边的类型而导致学习的节点embedding不准确。针对上述问题,提出一种在异构图上把边类型融入到元路径的方法,使节点在每个场景下得到单独训练。然后,使用图注意力机制将不同场景下的同一节点的embedding进行聚合,最终得到该节点的embedding。并用来预测用户与未交互的物品之间的行为关系,从而达到向用户推荐物品的目的。实验表明,提出的算法在三个公开数据集上都取得了性能提升,在阿里天池赛数据集上[F1]、ROC-AUC和PR-AUC指标分别提高了8.75%、6.03%和4.86%。

关键词: 分布式表示, 异构图, 元路径, 推荐系统