第17卷第4期 智能系统学报 Vol.17 No.4 2022年7月 CAAI Transactions on Intelligent Systems Jul.2022 D0:10.11992/tis.202107031 网络出版地址：https:/kns.cnki.net/kcms/detail/23.1538.TP.20220420.1339.002.html 融合社交关系的轻量级图卷积协同过滤推荐方法 朱金侠，孟祥福，邢长征，孙德伟，薛琪，关钧渤 (辽宁工程技术大学电子与信息工程学院，辽宁葫芦岛125105) 摘要：图卷积网络(graph convolution network,GCN)因其强大的建模能力得到了迅速发展，目前大部分研究工 作直接继承了GCN的复杂设计（如特征变换，非线性激活等），缺乏简化工作。另外，数据稀疏性和隐式负反 馈没有被充分利用，也是当前推荐算法的局限。为了应对以上问题，提出了一种融合社交关系的轻量级图卷积 协同过滤推荐模型。模型摒弃了GCN中特征变换和非线性激活的设计；利用社交关系从隐式负反馈中产生一 系列的中间反馈，提高了隐式负反馈的利用率；最后，通过双层注意力机制分别突出了邻居节点的贡献值和每 一层图卷积层学习向量的重要性。在2个公开的数据集上进行实验，结果表明所提模型的推荐效果优于当前 的图卷积协同过滤算法。 关键词：协同过滤；图卷积网络：注意力机制：社交关系；推荐系统：隐式负反馈；图嵌入；用户偏好 中图分类号：TP311文献标志码：A文章编号：1673-4785(2022)04-0788-10 中文引用格式：朱金侠，孟样福，邢长征，等.融合社交关系的轻量级图卷积协同过滤推荐方法.智能系统学报，2022,17(4)： 788-797. 英文引用格式：ZHU Jinxia,MENG Xiangfu,.XING Changzheng,etal.Light graph convolutional collaborative filtering recom- mendation approach incorporating social relationships J.CAAI transactions on intelligent systems,2022,17(4):788-797. Light graph convolutional collaborative filtering recommendation approach incorporating social relationships ZHU Jinxia,MENG Xiangfu,XING Changzheng,SUN Dewei,XUE Qi,GUAN Junbo (School of Electronics and Information Engineering,Liaoning Technical University,Huludao 125105,China) Abstract:Graph convolutional network (GCN)has rapidly developed due to their powerful modeling capability. However,much of the research up to now has directly inherited the complex design of GCN(e.g.,feature transforma- tion and nonlinear activation),which lacks thorough ablation analysis on GCN.Additionally,implicit feedback is not fully utilized,and data sparsity is not well resolved,which are also shortcomings of current recommendation algorithms. This paper proposes a light graph convolutional collaborative filtering recommendation approach that incorporates so- cial relationships to address such problems (F-LightGCCF).In GCN,the model abandons the design of feature trans- formation and nonlinear activation.Then it can generate a series of intermediate feedback from users'implicit negative feedback by taking advantage of social networking,improving the utilization of implicit negative feedback.Lastly,the importance of the contribution values of neighboring nodes and the learning vectors of each layer of the graph convolu- tion layer are aggregated separately using the dual attention mechanism.By conducting experiments on two publicly available datasets,the results show that the proposed model outperforms current graph convolutional collaborative filter- ing algorithms in the recommendation. Keywords:collaborative filtering;graph convolution network;attention mechanism;social relationships;recommenda- tion system;implicit negative feedback;graph embedding,user preference 移动网络的普遍应用，为人们带来选择麻痹 的困扰，推荐系统是解决信息过载问题的关键技 术。推荐的重点在于推测用户偏好和拓展用户视 收稿日期：2021-07-17.网络出版日期：2022-04-21. 基金项目：国家重点研发计划项目(2018YFB1402901):国家自 野。推荐的核心在于预测用户是否会与某个项目 然科学基金项目(61772249)：辽宁省教育厅一般项 目(LJ2019QL017). 进行交互，例如点击、评级、购买以及其他形式的 通信作者：孟祥福.E-mail:marxi(@I26.com 交互。现有的推荐方法凶大多单一利用可以直DOI: 10.11992/tis.202107031 网络出版地址: https://kns.cnki.net/kcms/detail/23.1538.TP.20220420.1339.002.html 融合社交关系的轻量级图卷积协同过滤推荐方法 朱金侠，孟祥福，邢长征，孙德伟，薛琪，关钧渤 （辽宁工程技术大学 电子与信息工程学院，辽宁 葫芦岛 125105） 摘 要：图卷积网络 (graph convolution network, GCN) 因其强大的建模能力得到了迅速发展，目前大部分研究工 作直接继承了 GCN 的复杂设计（如特征变换，非线性激活等），缺乏简化工作。另外，数据稀疏性和隐式负反 馈没有被充分利用，也是当前推荐算法的局限。为了应对以上问题，提出了一种融合社交关系的轻量级图卷积 协同过滤推荐模型。模型摒弃了 GCN 中特征变换和非线性激活的设计；利用社交关系从隐式负反馈中产生一 系列的中间反馈，提高了隐式负反馈的利用率；最后，通过双层注意力机制分别突出了邻居节点的贡献值和每 一层图卷积层学习向量的重要性。在 2 个公开的数据集上进行实验，结果表明所提模型的推荐效果优于当前 的图卷积协同过滤算法。 关键词：协同过滤；图卷积网络；注意力机制；社交关系；推荐系统；隐式负反馈；图嵌入；用户偏好 中图分类号：TP311 文献标志码：A 文章编号：1673−4785(2022)04−0788−10 中文引用格式：朱金侠, 孟祥福, 邢长征, 等. 融合社交关系的轻量级图卷积协同过滤推荐方法 [J]. 智能系统学报, 2022, 17(4): 788–797. 英文引用格式：ZHU Jinxia, MENG Xiangfu, XING Changzheng, et al. Light graph convolutional collaborative filtering recom￾mendation approach incorporating social relationships[J]. CAAI transactions on intelligent systems, 2022, 17(4): 788–797. Light graph convolutional collaborative filtering recommendation approach incorporating social relationships ZHU Jinxia，MENG Xiangfu，XING Changzheng，SUN Dewei，XUE Qi，GUAN Junbo (School of Electronics and Information Engineering, Liaoning Technical University, Huludao 125105, China) Abstract: Graph convolutional network (GCN) has rapidly developed due to their powerful modeling capability. However, much of the research up to now has directly inherited the complex design of GCN (e.g., feature transforma￾tion and nonlinear activation), which lacks thorough ablation analysis on GCN. Additionally, implicit feedback is not fully utilized, and data sparsity is not well resolved, which are also shortcomings of current recommendation algorithms. This paper proposes a light graph convolutional collaborative filtering recommendation approach that incorporates so￾cial relationships to address such problems (F-LightGCCF). In GCN, the model abandons the design of feature trans￾formation and nonlinear activation. Then it can generate a series of intermediate feedback from users’ implicit negative feedback by taking advantage of social networking, improving the utilization of implicit negative feedback. Lastly, the importance of the contribution values of neighboring nodes and the learning vectors of each layer of the graph convolu￾tion layer are aggregated separately using the dual attention mechanism. By conducting experiments on two publicly available datasets, the results show that the proposed model outperforms current graph convolutional collaborative filter￾ing algorithms in the recommendation. Keywords: collaborative filtering; graph convolution network; attention mechanism; social relationships; recommenda￾tion system; implicit negative feedback; graph embedding; user preference 移动网络的普遍应用，为人们带来选择麻痹 的困扰，推荐系统是解决信息过载问题的关键技 术。推荐的重点在于推测用户偏好和拓展用户视 野。推荐的核心在于预测用户是否会与某个项目 进行交互，例如点击、评级、购买以及其他形式的 交互。现有的推荐方法[1-2] 大多单一利用可以直 收稿日期：2021−07−17. 网络出版日期：2022−04−21. 基金项目：国家重点研发计划项目（2018YFB1402901）；国家自 然科学基金项目（61772249）；辽宁省教育厅一般项 目（LJ2019QL017）. 通信作者：孟祥福. E-mail：marxi@126.com. 第 17 卷第 4 期 智 能 系 统 学 报 Vol.17 No.4 2022 年 7 月 CAAI Transactions on Intelligent Systems Jul. 2022