作者投稿和查稿 主编审稿 专家审稿 编委审稿 远程编辑

计算机工程 ›› 2023, Vol. 49 ›› Issue (2): 70-80. doi: 10.19678/j.issn.1000-3428.0063844

• 人工智能与模式识别 • 上一篇    下一篇

知识增强的图神经网络序列推荐模型

李盼1,2, 解庆1,2, 李琳1, 刘永坚1,2   

  1. 1. 武汉理工大学 计算机与人工智能学院, 武汉 430070;
    2. 武汉理工大学 重庆研究院, 重庆 401135
  • 收稿日期:2022-01-26 修回日期:2022-03-17 发布日期:2022-07-19
  • 作者简介:李盼(1997-),女,硕士研究生,主研方向为推荐算法;解庆(通信作者),副教授、博士;李琳,教授、博士;刘永坚,教授。
  • 基金资助:
    重庆市自然科学基金(cstc2021jcyj-msxmX1013);湖北省重点研发计划项目(2021BAA030)。

Knowledge-Enhanced Graph Neural Network Model for Sequential Recommendation

LI Pan1,2, XIE Qing1,2, LI Lin1, LIU Yongjian1,2   

  1. 1. School of Computer Science and Artificial Intelligence, Wuhan University of Technology, Wuhan 430070, China;
    2. Chongqing Research Institute, Wuhan University of Technology, Chongqing 401135, China
  • Received:2022-01-26 Revised:2022-03-17 Published:2022-07-19

摘要: 现有基于图神经网络的序列推荐模型大多仅关注用户与项目交互的结构性信息,序列偏好的学习仅涉及项目交互顺序,缺乏项目自身的内容信息,并且未有效利用用户信息及挖掘项目之间更深层的语义关系。提出一种知识增强的图神经网络序列推荐模型KGGNN,引入知识图谱,并结合用户交互数据构建协同知识图谱,学习得到项目语义关联辅助信息以及用户关联辅助信息。将交互序列构建成有向序列图,利用门控图神经网络以及用户关联辅助信息学习序列中项目节点的结构性信息。通过注意力机制组合项目向量作为全局序列偏好,将最近交互的项目作为当前兴趣偏好,融合两者形成最终序列偏好,并结合项目语义关联辅助信息进行模型预测。在Amazon-Book、Last-FM、Yelp2018这3个公开数据集上的实验结果表明,辅助信息能有效提升序列推荐的准确性,该模型在命中率(HIT@K)和归一化折损累计增益(NDCG@K)2个指标上相较于GRU4Rec、NARM、SASRec等模型均有显著提升。当评估指标K值选取10时,与KGSR模型相比,其HIT@10指标在3个数据集上分别提升12.9%、4.5%、6.9%,NDCG@10指标在3个数据集上分别提升29.4%、5.7%、16.7%。

关键词: 序列推荐, 图神经网络, 知识图谱, 推荐系统, 用户行为, 深度学习

Abstract: Existing sequential recommendation models based on the Graph Neural Network (GNN) focus on the structural information of users' interaction with items, and the learning of sequence preference involves the sequence of interactive items, which lack content information and user information, and does not mine a deep semantic relationship between items.A knowledge-enhanced GNN sequence recommendation model(KGGNN) is proposed.First, a knowledge graph is introduced and combined with users' interaction data to construct a collaborative knowledge graph, which can obtain the relevant auxiliary information of items and users.By transforming an interaction sequence into a directed sequence graph, the algorithm uses the Gated Graph Neural Network (GGNN) and relevant auxiliary information of the users to learn the structural information of the item node in the sequence.The model combines the vectors of the items as the global sequence preference using the attention mechanism, and the recently interactive item is considered as the current interest preference.The two preferences are fused to form the final sequence preference, which is used with the semantic relevant auxiliary information of the current item to predict the next item.The experimental results of Amazon Book, Last FM, and Yelp 2018 indicate that the auxiliary information can effectively help improve the accuracy of the hit rate(HIT@K) and normalized discounted cumulative gain(NDCG@K) metrics.Compared with GRU4Rec, NARM, SASRec, and other models, the aforementioned indicators have significantly improved.When the K value of the evaluation index is 10, compared with the KGSR model, HIT@10 is improved by 12.9%, 4.5%, and 6.9%, and NDCG@10 is improved by 29.4%, 5.7%, and 16.7%.

Key words: sequential recommendation, Graph Neural Network(GNN), Knowledge Graph(KG), recommendation system, user behavior, deep learning

中图分类号: