[1]沈鑫科,李 勇,陈建伟,等.融合协同知识图谱和图卷积网络的推荐算法[J].计算机技术与发展,2024,34(01):150-157.[doi:10. 3969 / j. issn. 1673-629X. 2024. 01. 022]
　SHEN Xin-ke,LI Yong,CHEN Jian-wei,et al.Collaborative Knowledge Graph and Graph Convolution Network Based Recommendation Algorithm[J].,2024,34(01):150-157.[doi:10. 3969 / j. issn. 1673-629X. 2024. 01. 022]

点击复制

融合协同知识图谱和图卷积网络的推荐算法()

分享到：

《计算机技术与发展》[ISSN:1006-6977/CN:61-1281/TN]

卷:

34

期数:

2024年01期

页码:

150-157

栏目:

人工智能

出版日期:

2024-01-10

文章信息/Info

Title:

Collaborative Knowledge Graph and Graph Convolution Network Based Recommendation Algorithm

文章编号:

1673-629X(2024)01-0150-08

作者:

沈鑫科¹; 2 ; 李 勇¹; 2 ; 陈建伟² ; 陈囿任¹

1. 新疆师范大学 计算机科学技术学院,新疆 乌鲁木齐 830054;

2. 新疆电子研究所,新疆 乌鲁木齐 830013

Author(s):

SHEN Xin-ke1; 2 ; LI Yong1; 2 ; CHEN Jian-wei2 ; CHEN You-ren1

1. School of Computer Science and Technology,Xinjiang Normal University,Urumqi 830054,China;

2. Xinjiang Electronic Research Institute,Urumqi 830013,China

关键词:

推荐算法; 协同知识图谱; 注意力机制; 图卷积网络; 实体特征

Keywords:

recommendation algorithm; collaborative knowledge graph; attention mechanism; graph convolution network; entity feature

分类号:

TP391. 1

DOI:

10. 3969 / j. issn. 1673-629X. 2024. 01. 022

摘要:

推荐系统广泛应用于互联网,缓解信息过载问题。 现有研究通常将知识图谱引入推荐算法中,但不能有效获取用户与项目的高阶建模以及存在数据稀疏性的问题。 该文提出了一种融合协同知识图谱和图卷积网络的推荐算法(CKGCN) 。 首先,将用户-项目交互矩阵与项目知识图谱构建为协同知识图谱,利用知识感知注意力机制对邻居节点进行权重分配,递归地捕获用户和项目的特征向量,搜索用户对项目的潜在喜好,有效缓解数据稀疏性的问题。 其次,采用基于图卷积网络的邻域聚合算法捕捉每层实体网络之间的高阶联系,将实体与邻域实体聚合,丰富实体语义表示。 另外,通过交叉压缩单元协作处理项目特征向量与实体特征向量,探索二者的高阶特征交互,从而过滤实体的冗余信息、挖掘项目更深层次的联系。 最后,对用户特征向量与项目特征向量进行计算得出用户对项目的预测概率。 经过点击率预测及Top-k 推荐实验证明,在书籍 Book_Crossing 和音乐 Last. FM 两个公开的数据集上,该算法与五种基线算法相比较,AUC,ACC,F1,Recall@ k 和 Precision@ k 评价指标值均有提升,表明该模型具有良好的推荐性能。

Abstract:

The recommendation system is widely used in the Internet to alleviate the problem of information overload. The existingresearch usually introduces knowledge graph into recommendation algorithm,but it cannot effectively obtain the high-level modeling ofusers and projects and has the problem of data sparsity. We propose a collaborative knowledge graph and graph convolution networkbased recommendation algorithm ( CKGCN) . Firstly,the user project interaction matrix and the project knowledge graph are constructedas?

a collaborative knowledge graph. The weight of neighbor nodes is allocated using the knowledge awareness attention mechanism,thefeature vectors of users and projects are captured recursively,and the potential preferences of users for projects are searched to effectivelyalleviate the problem of data sparsity. Secondly,the neighborhood aggregation algorithm based on graph convolution network is used tocapture the higher-order relationship between each layer of entity network,aggregate entities and neighborhood entities,

and enrich entitysemantic representation. In addition, the cross - compression unit cooperatively processes the project feature vector and entity featurevector to explore their higher - order feature interaction, so as to filter the redundant information of entities and mine the deeperrelationship of projects. Finally,the user feature vector and the project feature vector are calculated to obtain the prediction probability ofthe user to the project. According to the hit rate prediction and Top-k recommendation experiment,on the two public datasets of Crossingand Music Last. FM,this model is compared with five baseline models,namely,AUC,ACC,F1 Recall@ k and Precision@ k,and the evaluation index values have been improved,indicating that the model has good recommendation performance.

常用功能