计算机科学与探索

• 学术研究 •    

双通道异构图神经网络序列推荐算法

邬锦琛,杨兴耀,于炯,李梓杨,黄擅杭,孙鑫杰   

  1. 新疆大学 软件学院,乌鲁木齐 830008

Dual channel heterogenous graph neural Network for Sequence Recommendation

WU Jinchen, YANG Xingyao, YU Jiong, LI Ziyang, HUANG Shanhang, SUN Xinjie   

  1. School of Software, Xingjiang University, Wulumuqi 830008, China

摘要: 基于用户行为序列的推荐系统的目的是根据上一次序列的顺序预测用户的下一次点击。目前的研究一般是根据用户行为序列中项目的转换来了解用户偏好。然而,行为序列中的其他有效信息被忽略,如用户配置文件,这会导致模型无法了解用户的特定偏好。本文提出了一种基于双通道异构图神经网络的用户行为序列推荐算法(Dual Channel Heterogenous Graph Neural Network,DC-HetGNN),该方法通过异构图神经网络通道和异构图线图通道学习行为序列嵌入,并捕获用户的特定偏好。DC-HetGNN会根据行为序列构造包含各种类型节点的异构图,可以捕获项目、用户和序列之间的依赖关系。其次,异构图神经网络通道和异构图线图通道捕获物品复杂转换及序列之间的交互信息,并学习包含用户信息的物品嵌入。最后,考虑到用户长期和短期偏好的影响,将局部和全局序列嵌入与注意力网络相结合,得到最终的序列嵌入。在两个电商用户行为序列数据集Diginetica和Tmall上进行的实验表明,DC-HetGNN与新近模型FGNN相比在指标平均倒数排名(Mean Reciprocal Rank,MRR)和召回率(Recall)中平均分别提升2.08%和0.78%,与TGSRec相比在指标MRR@n和Recall@n中平均分别提升2.70%和0.49%。

关键词: 推荐系统, 用户行为序列, 异构图神经网络

Abstract: The purpose of recommendation system based on user behavior sequence is to predict user's next click according to the order of last sequence. The current research is generally based on the conversion of items in the user behavior sequence to understand user preferences. However, other valid information in the behavior sequence, is ignored, such as the user profile, which results in the model failing to understand the user's specific preferences.  In this paper, a dual channel heterogenous graph neural network for user behaviour sequence recommendation (DC-HetGNN) is proposed. The method uses an heterogeneous graph neural network channel and a heterogeneous graph line channel to learn behavior sequence embedding and capture the specific preferences of users. DC-HetGNN constructs heterogeneous graphs containing various types of nodes based on behavior sequences that capture dependencies between projects, users, and sequences. Secondly, the heterogeneous graph neural network channel and the heterogeneous graph line channel capture the complex transformation of items and the interaction between the sequences, and learn the embedding of items containing user information. Finally, considering the influence of users' long-term and short-term preferences, local and global sequence embedding is combined with attention network to obtain the final sequence embedding. A large number of experiments conducted on Diginetica and Tmall, two real e-commerce user behavior sequence datasets, show that compared with recent model FGNN, DC-HETGNN improved by 2.08% and 0.78% on average in performance criterions Mean Reciprocal Rank and Recall, respectively, and by 2.70% and 0.49% in performance criterions MRR@n and Recall@n, respectively, compared with recent model TGSRec.

Key words: recommender system, user behaviour sequence, heterogenous graph neural network