基于知识图谱的双端邻居信息融合推荐算法

doi:10.3778/j.issn.1673-9418.2012054

计算机科学与探索 ›› 2022, Vol. 16 ›› Issue (6): 1354-1361.DOI: 10.3778/j.issn.1673-9418.2012054

基于知识图谱的双端邻居信息融合推荐算法

王宝亮¹^,⁺(), 潘文采²

1. 天津大学信息与网络中心,天津 300072
2. 天津大学电气自动化与信息工程学院,天津 300072

收稿日期:2020-12-15 修回日期:2021-04-02 出版日期:2022-06-01 发布日期:2021-04-15
通讯作者: + E-mail: wbl@tju.edu.cn
作者简介:王宝亮（1971—）,男,山东潍坊人,博士,高级工程师,硕士生导师,主要研究方向为数据挖掘、移动互联、图像处理等。
潘文采（1995—）,男,安徽黄山人,硕士研究生,主要研究方向为推荐算法。
基金资助:
赛尔网络下一代互联网技术创新项目(NGII20170104)

Two-Terminal Neighbor Information Fusion Recommendation Algorithm Based on Knowledge Graph

WANG Baoliang¹^,⁺(), PAN Wencai²

1. Information and Network Center, Tianjin University, Tianjin 300072, China
2. School of Electrical Automation and Information Engineering, Tianjin University, Tianjin 300072, China

Received:2020-12-15 Revised:2021-04-02 Online:2022-06-01 Published:2021-04-15
About author:WANG Baoliang, born in 1971, Ph.D., senior engineer, M.S. supervisor. His research interests include data mining, mobile Internet, image processing, etc.
PAN Wencai, born in 1995, M.S. candidate. His research interest is recommendation algorithm.
Supported by:
CERNET Innovation Project(NGII20170104)

摘要/Abstract

摘要：

针对一些基于知识图谱的推荐算法仅聚合一端邻居而无法有效确定实体与用户之间关系的问题,提出了一种基于知识图谱的双端邻居聚合推荐算法,算法通过探究知识图谱的内在联系以发掘用户和物品之间的潜在关系。在用户端,提出了一种聚合用户邻居信息的方法,在知识图谱的关系空间下,使用知识图谱来传播和提取用户的潜在兴趣,通过迭代将潜在兴趣注入具有注意偏差的用户特征中生成用户嵌入表示向量;在物品端,将融合了用户邻居信息的用户向量送入KGCN模型,并在聚合物品和其邻居信息时,采用新的聚合方式,生成物品嵌入表示向量。最后,将得到的用户和物品向量送入预测环节,通过向量的内积运算并归一化得到用户和物品的关联分数,然后在训练集中进行训练,优化参数。在两个公开数据集上进行对比实验,在 Book-Crossing数据集上,相较于最优基线,AUC和ACC分别提升了1.72%和4.24%;在Last.FM数据集上,AUC和ACC分别提升了1.07%和1.14%,从而验证在聚合两端邻居信息后,算法的有效性得到了提升。

关键词: 推荐算法, 知识图谱, 邻居聚合

Abstract:

In view of the problem that some recommendation algorithms based on knowledge graphs only aggregate one end of the neighbors and cannot effectively determine the relationship between entities and users, this paper proposes a dual-end neighbor aggregation recommendation algorithm based on knowledge graphs. This algorithm explores the internal connections of knowledge graphs to discover the potential relationship between users and items. On the user side, this paper proposes a method of aggregating user neighbor information. In the relational space of the knowledge graph, the knowledge graph is used to spread and extract the user’s potential interest, and iteratively inject the potential interest into the user characteristics with attention bias to generate user embedding representation vector. At the item side, the user vector that aggregates the user’s neighbor information is sent to the KGCN (knowledge graph convolutional networks) model, and when the polymer product and its neighbor infor-mation are used, a new aggregation method is used to generate the item embedding representation. Finally, the obtained vector is sent to train. Through the inner product operation of the vector and normalization, the association score between the user and the item is obtained. Then the training is carried out in the training set to optimize the parameters. Comparative experiments are conducted on two public datasets. Compared with the baseline, on the Book-Crossing dataset, AUC and ACC are increased by 1.72% and 4.24%, and on the Last.FM dataset, AUC and ACC are increased by 1.07% and 1.14%. It is proven that the effectiveness of the algorithm is improved after the information of neighbors at both ends is aggregated.

Key words: recommendation algorithm, knowledge graph, neighbor aggregation

中图分类号:

TP391

王宝亮, 潘文采. 基于知识图谱的双端邻居信息融合推荐算法[J]. 计算机科学与探索, 2022, 16(6): 1354-1361.

WANG Baoliang, PAN Wencai. Two-Terminal Neighbor Information Fusion Recommendation Algorithm Based on Knowledge Graph[J]. Journal of Frontiers of Computer Science and Technology, 2022, 16(6): 1354-1361.

图/表 12

图1 总体框架

Fig.1 Overall framework

图2 用户节点信息融合

Fig.2 User node information aggregation

图3 KGCN模型

Fig.3 KGCN model

表1 两个数据集的统计数据

Table 1 Statistics for two datasets

统计项目	Book-Crossing	Last.FM
users	19 676	1 872
items	20 003	3 846
interactions	172 576	42 346
entities	25 787	9 366
relations	18	60
KG triples	60 787	15 518

表2 参数的设置

Table 2 Parameter setting

Parameter	Book-Crossing	Last.FM
$K - u$	16	64
$K - i$	3	3
$H - u$	3	3
$H - i$	1	1
$d$	32	16
$λ$	1×10^-6	1×10^-6
$η$	1×10^-4	5×10^-3
$γ$	0.1	0.1
Batch size	256	128

表2 参数的设置

Table 2 Parameter setting

Parameter	Book-Crossing	Last.FM
$K - u$	16	64
$K - i$	3	3
$H - u$	3	3
$H - i$	1	1
$d$	32	16
$λ$	1×10^-6	1×10^-6
$η$	1×10^-4	5×10^-3
$γ$	0.1	0.1
Batch size	256	128

表3 CTR预测中AUC和ACC的结果

Table 3 Results of AUC and ACC in CTR prediction

模型	Book-Crossing		Last.FM
模型	AUC	ACC	AUC	ACC
LibFM	0.685 0	0.640 0	0.777 0	0.709 0
Wide&Deep	0.712 0	0.624 0	0.756 0	0.688 0
RippleNet	0.724 9	0.662 0	0.813 2	0.746 0
KGCN	0.691 7	0.631 8	0.803 5	0.732 2
KGNN-LS	0.687 3	0.632 0	0.802 3	0.728 2
本文模型	0.737 4	0.690 1	0.821 9	0.754 5
本文模型-u	0.728 0	0.664 1	0.817 6	0.743 9
本文模型-i	0.673 8	0.619 9	0.773 0	0.700 8
不同预测函数	0.733 0	0.693 9	0.792 0	0.753 6

表4 参数 γ对AUC值的影响

Table 4 Influence of parameter γ on AUC value

$γ$	Book-Crossing	Last.FM
100	0.734 9	0.814 1
10	0.737 1	0.820 6
1	0.736 1	0.820 4
0.1	0.737 4	0.821 9
0.01	0.735 8	0.821 1
直接累加	0.737 2	0.818 5

表4 参数 γ对AUC值的影响

Table 4 Influence of parameter γ on AUC value

$γ$	Book-Crossing	Last.FM
100	0.734 9	0.814 1
10	0.737 1	0.820 6
1	0.736 1	0.820 4
0.1	0.737 4	0.821 9
0.01	0.735 8	0.821 1
直接累加	0.737 2	0.818 5

表5 用户端每跳采样值 K - u对AUC值的影响

Table 5 Impact of K - u sampled value per hop on AUC value at user end

$K - u$	Book-Crossing	Last.FM
2	0.725 5	0.806 3
4	0.729 7	0.807 2
8	0.735 4	0.813 5
16	0.737 4	0.819 0
32	0.736 4	0.819 6
64	0.735 5	0.821 9

表5 用户端每跳采样值 K - u对AUC值的影响

Table 5 Impact of K - u sampled value per hop on AUC value at user end

$K - u$	Book-Crossing	Last.FM
2	0.725 5	0.806 3
4	0.729 7	0.807 2
8	0.735 4	0.813 5
16	0.737 4	0.819 0
32	0.736 4	0.819 6
64	0.735 5	0.821 9

表6 物品端每跳采样值 K - i对ACC值的影响

Table 6 Impact of K - i sampled value per hop on ACC value at item end

$K - i$	Book-Crossing	Last.FM
1	0.670 2	0.749 2
2	0.680 0	0.741 6
3	0.690 1	0.754 5
4	0.689 4	0.739 7
5	0.691 3	0.752 6
6	0.691 4	0.746 8
8	0.689 6	0.745 6

表6 物品端每跳采样值 K - i对ACC值的影响

Table 6 Impact of K - i sampled value per hop on ACC value at item end

$K - i$	Book-Crossing	Last.FM
1	0.670 2	0.749 2
2	0.680 0	0.741 6
3	0.690 1	0.754 5
4	0.689 4	0.739 7
5	0.691 3	0.752 6
6	0.691 4	0.746 8
8	0.689 6	0.745 6

表7 用户端聚合跳数 H - u对AUC值的影响

Table 7 Impact of user end aggregation hops H - u on AUC value

$H - u$	Book-Crossing	Last.FM
1	0.733 3	0.806 2
2	0.734 1	0.808 0
3	0.737 4	0.821 9
4	0.735 7	0.811 4

表7 用户端聚合跳数 H - u对AUC值的影响

Table 7 Impact of user end aggregation hops H - u on AUC value

$H - u$	Book-Crossing	Last.FM
1	0.733 3	0.806 2
2	0.734 1	0.808 0
3	0.737 4	0.821 9
4	0.735 7	0.811 4

表8 物品端聚合跳数 H - i对ACC值的影响

Table 8 Impact of item end aggregation hops H - i on ACC value

$H - i$	Book-Crossing	Last.FM
1	0.690 1	0.754 5
2	0.444 1	0.733 3
3	0.445 0	0.734 3
4	0.447 2	0.732 8

表8 物品端聚合跳数 H - i对ACC值的影响

Table 8 Impact of item end aggregation hops H - i on ACC value

$H - i$	Book-Crossing	Last.FM
1	0.690 1	0.754 5
2	0.444 1	0.733 3
3	0.445 0	0.734 3
4	0.447 2	0.732 8

表9 向量维度值 d对AUC值的影响

Table 9 Influence of vector dimension value d on AUC value

$d$	Book-Crossing	Last.FM
4	0.739 1	0.806 2
8	0.736 7	0.819 5
16	0.735 5	0.821 9
32	0.737 4	0.820 6
64	0.736 3	0.814 5
128	0.734 4	0.815 9

表9 向量维度值 d对AUC值的影响

Table 9 Influence of vector dimension value d on AUC value

$d$	Book-Crossing	Last.FM
4	0.739 1	0.806 2
8	0.736 7	0.819 5
16	0.735 5	0.821 9
32	0.737 4	0.820 6
64	0.736 3	0.814 5
128	0.734 4	0.815 9

参考文献 15

[1]	KOREN Y, BELL R, VOLINSKY C. Matrix factorization techniques for recommender systems[J]. Computer, 2009, 42(8): 30-37.
[2]	SHI Y, LARSON M, HANJALIC A. Collaborative filtering beyond the user-item matrix[J]. ACM Computing Surveys, 2014, 47(1): 1-45.
[3]	MOHSEN J, MARTIN E. A matrix factorization technique with trust propagation for recommendation in social networks[C]// Proceedings of the 4th ACM Conference on Recom-mender Systems, Barcelona, Spain, Sep 26-30, 2010. New York: ACM, 2010: 135-142.
[4]	WANG H, WANG J, ZHAO M, et al. Joint topic-semantic-aware social recommendation for online voting[C]// Procee-dings of the 2017 ACM on Conference on Information and Knowledge Management, Singapore, Nov 6-10, 2017. New York: ACM, 2017: 347-356.
[5]	WANG H, ZHANG F, HOU M, et al. Shine: signed hetero-geneous information network embedding for sentiment link prediction[C]// Proceedings of the 11th ACM International Conference on Web Search and Data Mining, Los Angeles, Feb 5-9, 2018. New York: ACM, 2018: 592-600.
[6]	WANG Q, MAO Z, WANG B, et al. Knowledge graph em-bedding: a survey of approaches and applications[J]. IEEE Transactions on Knowledge and Data Engineering, 2017, 29(12): 2724-2743. DOI URL
[7]	WANG H, ZHANG F, XIE X, et al. DKN: deep knowledge-aware network for news recommendation[C]// Proceedings of the 2018 World Wide Web Conference, Lyon, Apr 23-27, 2018. New York: ACM, 2018: 1835-1844.
[8]	WANG H, ZHANG F, WANG J, et al. RippleNet: propagating user preferences on the knowledge graph for recommender systems[C]// Proceedings of the 27th ACM International Conference on Information and Knowledge Management, Torino, Oct 22-26, 2018. New York: ACM, 2018: 417-426.
[9]	WANG H, ZHANG F, WANG J, et al. Exploring high-order user preference on the knowledge graph for recommender systems[J]. ACM Transactions on Information Systems, 2019, 37(3): 1-26.
[10]	WANG H, ZHAO M, XIE X, et al. Knowledge graph convo-lutional networks for recommender systems[C]// Proceedings of the 2019 World Wide Web Conference, San Francisco, May 13-17, 2019. New York: ACM, 2019: 3307-3313.
[11]	WANG H, ZHANG F, ZHANG M, et al. Knowledge-aware graph neural networks with label smoothness regularization for recommender systems[C]// Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, Anchorage, Aug 4-8, 2019. New York: ACM, 2019: 968-977.
[12]	WANG X, WANG D, XU C, et al. Explainable reasoning over knowledge graphs for recommendation[C]// Proceedings of the 33rd AAAI Conference on Artificial Intelligence, the 31st Innovative Applications of Artificial Intelligence Con-ference, the 9th AAAI Symposium on Educational Adv-ances in Artificial Intelligence, Honolulu, Jan 27-Feb 1, 2019. Menlo Park: AAAI, 2019: 5329-5336.
[13]	WANG X, HE X N, CAO Y X, et al. KGAT: knowledge graph attention network for recommendation[C]// Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, Anchorage, Aug 4-8, 2019. New York: ACM, 2019: 950-958.
[14]	RENDLE S. Factorization machines with LibFM[J]. ACM Transactions on Intelligent Systems and Technology, 2012, 3(3): 57.
[15]	CHENG H T, KOC L, HARMSEN J, et al. Wide & deep learning for recommender systems[C]// Proceedings of the 1st Workshop on Deep Learning for Recommender Systems, Boston, Sep 15, 2016. New York: ACM, 2016: 7-10.

基于知识图谱的双端邻居信息融合推荐算法

Two-Terminal Neighbor Information Fusion Recommendation Algorithm Based on Knowledge Graph

RichHTML

PDF

可视化

摘要/Abstract

引用本文

使用本文

图/表 12

参考文献 15

相关文章 15

编辑推荐

Metrics

[1]	于慧琳, 陈炜, 王琪, 高建伟, 万怀宇. 使用子图推理实现知识图谱关系预测[J]. 计算机科学与探索, 2022, 16(8): 1800-1808.
[2]	萨日娜, 李艳玲, 林民. 知识图谱推理问答研究综述[J]. 计算机科学与探索, 2022, 16(8): 1727-1741.
[3]	田萱, 陈杭雪. 推荐任务中知识图谱嵌入应用研究综述[J]. 计算机科学与探索, 2022, 16(8): 1681-1705.
[4]	韩毅, 乔林波, 李东升, 廖湘科. 知识增强型预训练语言模型综述[J]. 计算机科学与探索, 2022, 16(7): 1439-1461.
[5]	郭晓旺, 夏鸿斌, 刘渊. 融合知识图谱与图卷积网络的混合推荐模型[J]. 计算机科学与探索, 2022, 16(6): 1343-1353.
[6]	董文波, 孙仕亮, 殷敏智. 医学知识推理研究现状与发展[J]. 计算机科学与探索, 2022, 16(6): 1193-1213.
[7]	张子辰, 岳昆, 祁志卫, 段亮. 时序知识图谱的增量构建[J]. 计算机科学与探索, 2022, 16(3): 598-607.
[8]	李想, 杨兴耀, 于炯, 钱育蓉, 郑捷. 基于知识图谱卷积网络的双端推荐算法[J]. 计算机科学与探索, 2022, 16(1): 176-184.
[9]	吴正洋, 汤庸, 刘海. 个性化学习推荐研究综述[J]. 计算机科学与探索, 2022, 16(1): 21-40.
[10]	武家伟, 孙艳春. 融合知识图谱和深度学习方法的问诊推荐系统[J]. 计算机科学与探索, 2021, 15(8): 1432-1440.
[11]	高仰, 刘渊. 融合知识图谱和短期偏好的推荐算法[J]. 计算机科学与探索, 2021, 15(6): 1133-1144.
[12]	舒世泰，李松，郝晓红，张丽平. 知识图谱嵌入技术研究进展[J]. 计算机科学与探索, 2021, 15(11): 2048-2062.
[13]	陈子睿, 王鑫, 王林, 徐大为, 贾勇哲. 开放领域知识图谱问答研究综述[J]. 计算机科学与探索, 2021, 15(10): 1843-1869.
[14]	蔺奇卡, 张玲玲, 刘均, 赵天哲. 基于问句感知图卷积的教育知识库问答方法[J]. 计算机科学与探索, 2021, 15(10): 1880-1887.
[15]	韩鑫鑫，贲可荣，张献. 军用软件测试领域的命名实体识别技术研究[J]. 计算机科学与探索, 2020, 14(5): 740-748.