基于协同知识嵌入注意网络的推荐算法研究
RECOMMENDATION ALGORITHM BASED ON COLLABORATIVE KGE-GUIDED ATTENTIVE NETWORK
师博雅 1梁光成 2孙宇健 2张家华 1胡泉2
作者信息
- 1. 中国航天系统科学与工程研究院 北京 100048
- 2. 航天宏康智能科技(北京)有限公司 北京 100048
- 折叠
摘要
推荐系统广泛用于实际应用场景中,现有利用图神经网络的基于知识图谱的解决方法忽略了对知识图谱中蕴含丰富关系知识的有效编码.提出一种协同知识嵌入注意网络模型,通过在知识高阶传播过程中融入翻译模型对以三元组形式表示的知识进行有效编码,实现协同信息与知识传播更高效的融合方式,并通过拓展端到端模型CKAN完成推荐算法的设计.在三个真实推荐场景中的实验结果表明,相对于现有的KGCN、KGNN-LS、KGAT等方法,所提算法在点击率预测中的AUC值在Last.FM数据集上达到了 0.847 3,在Book-Crossing数据集上达到了 0.753 8,在Dianping-Food数据集上达到了 0.878 3;在Top-K推荐中的召回率均优于基准算法.
Abstract
Recommendation systems are widely used in practical application scenarios,the existing solutions based on knowledge graphs using graph convolutional neural networks ignore effective embedding the rich semantic information contained in the knowledge graph(KG).This paper proposes a collaborative knowledge graph embedding(KGE)-guided attentive network.By incorporating a translational distance model into the knowledge high-order propagation,the proposed algorithm effectively embedded knowledge in triple sets,and a more efficient way of fusion collaborative information and auxiliary knowledge was realized.And by further expanding the end-to-end model CKAN,the design of the recommendation algorithm was completed.The test results in three real recommendation scenarios show that compared with the existing methods such as KGCN,KGNN-LS,and KGAT,the AUC of the proposed algorithm in CTR prediction reaches 0.847 3 on the Last.FM,reaches 0.753 8 on the Book-Crossing,reaches 0.878 3 on the Dianping-Food,and the recall value is better than benchmark algorithms in Top-K recommendation.
关键词
推荐系统/协同过滤/异质信息网络/知识图谱嵌入/图卷积神经网络Key words
Recommender system/Collaborative filtering/Heterogeneous information network/Knowledge graph embedding/Graph convolutional neural network引用本文复制引用
出版年
2024
国家科技期刊平台
NSTL
维普
万方数据