计算机科学与探索 ›› 2023, Vol. 17 ›› Issue (9): 2209-2218.DOI: 10.3778/j.issn.1673-9418.2204012

• 人工智能·模式识别 • 上一篇    下一篇

知识水波图卷积网络推荐模型

崔焕庆,宋玮情,杨峻铸   

  1. 1. 山东科技大学 计算机科学与工程学院,山东 青岛 266590
    2. 浪潮集团有限公司 高效能服务器和存储技术国家重点实验室,济南 250014
  • 出版日期:2023-09-01 发布日期:2023-09-01

Knowledge Ripple Graph Convolutional Network for Recommendation

CUI Huanqing, SONG Weiqing, YANG Junzhu   

  1. 1. College of Computer Science and Engineering, Shandong University of Science and Technology, Qingdao, Shandong 266590, China
    2. State Key Laboratory of High-End Server & Storage Technology, Inspur Group Co., Ltd., Jinan 250014, China
  • Online:2023-09-01 Published:2023-09-01

摘要: 为了解决基于知识图谱的推荐系统中常见的高阶建模难和用户特征建模不充分问题,提出了以端到端的方法获取知识图谱中高阶语义信息的知识水波图卷积网络(KRGCN)。模型分为用户特征学习和项目特征学习两部分,其中用户特征学习基于用户历史交互记录,利用偏好传播和交叉压缩单元进行嵌入传播来获得用户特征表示;项目特征学习是在项目知识图谱上利用图卷积网络聚合每一项目的高阶邻域信息,使用偏差区分项目邻域的重要性来获得项目特征表示。最后,利用用户特征表示和项目特征表示计算预测值,利用向量之间的内积操作计算用户与项目的交互概率。在两个公开数据集上,使用五个基线方法进行了对比实验。在Book-Crossing数据集上,KRGCN的AUC、ACC和F1值分别提升了4.43%~11.96%、1.68%~10.82%和1.90%~12.78%;在Last.FM数据集上,KRGCN的AUC、ACC和F1值分别提升了2.94%~16.84%、2.36%~16.59%和0.83%~17.69%。实验结果表明,KRGCN能够同时实现用户和项目的高阶建模,与其他代表性的模型相比有良好的推荐效果。

关键词: 推荐系统, 知识图谱, 偏好传播, 图卷积网络(GCN)

Abstract: This paper proposes a knowledge ripple graph convolutional network (KRGCN) to obtain high-order semantic information from knowledge graph in an end-to-end approach, which addresses the issues of high-order modeling and inadequate user feature modeling in the recommendation system based on knowledge graph. The model consists of user and item features learning. Based on the history of user interaction, the user feature learning applies the preference propagation and cross compression to embedded propagation to obtain user representation. The item feature learning applies graph convolutional network to aggregating the high-order neighborhood information of each item on the item knowledge graph, and uses the bias to distinguish the importance of the neighbors to obtain the item representation.Finally, the predicted values are calculated using user and item feature representation, and the interaction probability between user and item is calculated by inner product operation between vectors. On two public datasets, comparative experiments are conducted using five baseline methods. For the Book-Crossing dataset, the AUC, ACC and F1 values of KRGCN are respectively increased by 4.43% to 11.96%, 1.68% to 10.82% and 1.90% to 12.78%. For the Last.FM dataset, the AUC, ACC and F1 values of KRGCN are respectively increased by 2.94% to 16.84%, 2.36% to 16.59% and 0.83% to 17.69%. The results show that KRGCN can implement high-order modeling of user and item simultaneously, and it outperforms the other models in recommendation.

Key words: recommendation system, knowledge graph, preference propagation, graph convolutional network (GCN)