Computer Engineering and Applications ›› 2022, Vol. 58 ›› Issue (9): 107-115.DOI: 10.3778/j.issn.1002-8331.2105-0241

• Big Data and Cloud Computing • Previous Articles     Next Articles

Neighborhood Awareness Graph Neural Networks for Session-Based Recommendation

HE Qianqian, SUN Jingyu, ZENG Yazhu   

  1. College of Software, Taiyuan University of Technology, Jinzhong, Shanxi 030600, China
  • Online:2022-05-01 Published:2022-05-01

基于邻域感知图神经网络的会话推荐

何倩倩,孙静宇,曾亚竹   

  1. 太原理工大学 软件学院,山西 晋中 030600

Abstract: Graph neural network and its improved models show better performances in session-based recommendation.They convert session as graph structure, and capture item features from transformation relationship between items in graph structure. However, most of the models ignore that there may be useful informations between different sessions, which can support the prediction task. Therefore, it proposes neighborhood aware graph neural networks for session-based recommendation(NA-GNN). It builds session layer graph and global neighborhood layer graph, and captures item representations from them. Next it uses the attention mechanism to aggregate item representations. So then, it combines the maximum mutual information between sessions into network training. Experiments on two real datasets show that the prediction accuracy is better than others:P@20 increased by 1.85% on Yoochoose and 7.19% on Diginetica; MRR@20 increased by 0.48% and 8.36%, which proves the model is effective and reasonable.

Key words: neighborhood awareness, graph neural network, attention mechanism, session recommendation

摘要: 在基于会话的推荐中,图神经网络及其改进模型将会话内复杂的交互关系建模为图结构并从中捕获项目特征,是现有推荐模型中性能较好的一类方法。然而大多数模型都忽略了不同会话之间可能存在的有效信息,仅对当前会话建模难以利用其他会话,也无法发挥邻域信息的辅助作用。因此提出基于邻域感知图神经网络的会话推荐(NA-GNN)。该模型构建会话层和全局邻域层的图结构捕获项目表示,结合注意力机制聚合两种项目表征,将会话序列之间的互信息最大化地结合到网络训练中。在真实的数据集Yoochoose和Diginetica上进行实验,与性能最优的基准模型相比,模型P@20在Yoochoose上提高了1.85%,在Diginetica上提升了7.19%;MRR@20分别提升了0.48%和8.36%,证明模型的有效性和合理性。

关键词: 邻域感知, 图神经网络, 注意力机制, 会话推荐